File: | nnc/ccv_cnnp_model_addons.c |
Warning: | line 400, column 3 Declared variable-length array (VLA) has negative size |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | #include "ccv_nnc.h" | |||
2 | #include "ccv_nnc_easy.h" | |||
3 | #include "ccv_nnc_internal.h" | |||
4 | #include "ccv_internal.h" | |||
5 | #include "_ccv_cnnp_model.h" | |||
6 | ||||
7 | // MARK - Add-on Functions | |||
8 | ||||
9 | static int _ccv_cnnp_model_clip_grad_norm_reduce_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context) | |||
10 | { | |||
11 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8); | |||
12 | ccv_nnc_tensor_t* const old_norm2 = outputs[1 + device_id * 2]; | |||
13 | ccv_nnc_tensor_t* const norm2 = outputs[1 + device_id * 2 + 1]; | |||
14 | const int tensor_count = ccv_nnc_tensor_count(inputs[0]->info); | |||
15 | if (tensor_count == 1) | |||
16 | ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], inputs[0])(ccv_nnc_tensor_t* []){inputs[0], inputs[0]}, (1 +1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
17 | else { | |||
18 | ccv_nnc_cmd_exec(CMD_REDUCE_NORM2_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_NORM2_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0), hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
19 | ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(norm2, norm2)(ccv_nnc_tensor_t* []){norm2, norm2}, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
20 | } | |||
21 | ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1, 1}}}, 0), hint, flags, TENSOR_LIST(old_norm2, norm2)(ccv_nnc_tensor_t* []){old_norm2, norm2}, (1 +1 +1 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_norm2)(ccv_nnc_tensor_t* []){old_norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
22 | return CCV_NNC_EXEC_SUCCESS; | |||
23 | } | |||
24 | ||||
25 | static ccv_nnc_cmd_vtab_t clip_grad_norm_reduce_norm2_vtab = { | |||
26 | .exec = _ccv_cnnp_model_clip_grad_norm_reduce_norm2 | |||
27 | }; | |||
28 | ||||
29 | static int _ccv_cnnp_model_clip_grad_norm_scatter_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context) | |||
30 | { | |||
31 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8); | |||
32 | ccv_nnc_tensor_t* const norm2 = inputs[1 + device_id * 2]; | |||
33 | ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], norm2)(ccv_nnc_tensor_t* []){inputs[0], norm2}, (1 +1 +1 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(outputs[0])(ccv_nnc_tensor_t* []){outputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
34 | return CCV_NNC_EXEC_SUCCESS; | |||
35 | } | |||
36 | ||||
37 | static ccv_nnc_cmd_vtab_t clip_grad_norm_scatter_norm2_vtab = { | |||
38 | .exec = _ccv_cnnp_model_clip_grad_norm_scatter_norm2 | |||
39 | }; | |||
40 | ||||
41 | void ccv_cnnp_model_parameters_clip_grad_norm(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, int norm_type, float max_norm, ccv_nnc_stream_context_t* const stream_context) | |||
42 | { | |||
43 | assert(norm_type == 2)((void) sizeof ((norm_type == 2) ? 1 : 0), __extension__ ({ if (norm_type == 2) ; else __assert_fail ("norm_type == 2", "ccv_cnnp_model_addons.c" , 43, __extension__ __PRETTY_FUNCTION__); })); | |||
44 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; | |||
45 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c" , 45, __extension__ __PRETTY_FUNCTION__); })); | |||
46 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); | |||
47 | ccv_nnc_tensor_t* norm2[parallel_count * 2]; | |||
48 | ccv_nnc_tensor_t* max_normt[parallel_count]; | |||
49 | const int stream_type = model->compiled_data->stream_type; | |||
50 | int i; | |||
51 | if (stream_type == CCV_STREAM_CONTEXT_GPU) | |||
52 | { | |||
53 | for (i = 0; i < parallel_count; i++) | |||
54 | { | |||
55 | ccv_nnc_tensor_param_t info = { | |||
56 | .type = CCV_TENSOR_GPU_MEMORY, | |||
57 | .format = CCV_TENSOR_FORMAT_NHWC, | |||
58 | .datatype = CCV_32F, | |||
59 | .dim = {1}, | |||
60 | }; | |||
61 | CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff ) << 8)); | |||
62 | norm2[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0); | |||
63 | norm2[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0); | |||
64 | max_normt[i] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0); | |||
65 | } | |||
66 | } else { | |||
67 | for (i = 0; i < parallel_count; i++) | |||
68 | { | |||
69 | ccv_nnc_tensor_param_t info = { | |||
70 | .type = CCV_TENSOR_CPU_MEMORY, | |||
71 | .format = CCV_TENSOR_FORMAT_NHWC, | |||
72 | .datatype = CCV_32F, | |||
73 | .dim = {1}, | |||
74 | }; | |||
75 | norm2[i * 2] = ccv_nnc_tensor_new(0, info, 0); | |||
76 | norm2[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0); | |||
77 | max_normt[i] = ccv_nnc_tensor_new(0, info, 0); | |||
78 | } | |||
79 | } | |||
80 | // zero out old norm2. | |||
81 | if (parallel_count > 1) | |||
82 | { | |||
83 | ccv_nnc_stream_context_t* streams[parallel_count]; | |||
84 | ccv_nnc_stream_signal_t* signal; | |||
85 | if (stream_context) | |||
86 | signal = ccv_nnc_stream_context_emit_signal_new(stream_context); | |||
87 | for (i = 0; i < parallel_count; i++) | |||
88 | { | |||
89 | const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU; | |||
90 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8); | |||
91 | int type = stream_type; | |||
92 | CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff ) << 8)); | |||
93 | ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type); | |||
94 | // Wait signal to finish. | |||
95 | if (stream_context) | |||
96 | ccv_nnc_stream_context_wait_signal(stream_0, signal); | |||
97 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
98 | if (stream_context) | |||
99 | { | |||
100 | ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0); | |||
101 | ccv_nnc_stream_context_wait_signal(stream_context, signal); | |||
102 | } | |||
103 | streams[i] = stream_0; | |||
104 | } | |||
105 | // If this should be blocking, blocking it. | |||
106 | if (!stream_context) | |||
107 | for (i = 0; i < parallel_count; i++) | |||
108 | if (streams[i]) | |||
109 | ccv_nnc_stream_context_wait(streams[i]); | |||
110 | } else { | |||
111 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
112 | } | |||
113 | // Gather norm2. | |||
114 | ccv_nnc_cmd_t reduce_cmd = { | |||
115 | .cmd = CCV_NNC_CUSTOM_FORWARD, | |||
116 | .isa = &clip_grad_norm_reduce_norm2_vtab, | |||
117 | }; | |||
118 | ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, norm2, parallel_count * 2, stream_context); | |||
119 | // Now compute max(max_norm / norm2, 1.0). | |||
120 | if (parallel_count > 1) | |||
121 | { | |||
122 | ccv_nnc_stream_context_t* streams[parallel_count]; | |||
123 | ccv_nnc_stream_signal_t* signal; | |||
124 | if (stream_context) | |||
125 | signal = ccv_nnc_stream_context_emit_signal_new(stream_context); | |||
126 | for (i = 0; i < parallel_count; i++) | |||
127 | { | |||
128 | const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU; | |||
129 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8); | |||
130 | int type = stream_type; | |||
131 | CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff ) << 8)); | |||
132 | ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type); | |||
133 | // Wait signal to finish. | |||
134 | if (stream_context) | |||
135 | ccv_nnc_stream_context_wait_signal(stream_0, signal); | |||
136 | ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
137 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[i])(ccv_nnc_tensor_t* []){max_normt[i]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
138 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[i], norm2[i * 2])(ccv_nnc_tensor_t* []){max_normt[i], norm2[i * 2]}, (1 +1 +1 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1 ), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
139 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){. size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1 }}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
140 | if (stream_context) | |||
141 | { | |||
142 | ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0); | |||
143 | ccv_nnc_stream_context_wait_signal(stream_context, signal); | |||
144 | } | |||
145 | streams[i] = stream_0; | |||
146 | } | |||
147 | // If this should be blocking, blocking it. | |||
148 | if (!stream_context) | |||
149 | for (i = 0; i < parallel_count; i++) | |||
150 | if (streams[i]) | |||
151 | ccv_nnc_stream_context_wait(streams[i]); | |||
152 | } else { | |||
153 | ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
154 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[0])(ccv_nnc_tensor_t* []){max_normt[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
155 | ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[0], norm2[0])(ccv_nnc_tensor_t* []){max_normt[0], norm2[0]}, (1 +1 +1 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
156 | ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){. size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1 }}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
157 | } | |||
158 | ccv_nnc_cmd_t scatter_cmd = { | |||
159 | .cmd = CCV_NNC_CUSTOM_FORWARD, | |||
160 | .isa = &clip_grad_norm_scatter_norm2_vtab, | |||
161 | }; | |||
162 | ccv_cnnp_model_parameter_gradients_map(model, parameters, scatter_cmd, ccv_nnc_no_hint, 0, norm2, parallel_count * 2, 0, 0, stream_context); | |||
163 | if (stream_type == CCV_STREAM_CONTEXT_GPU) | |||
164 | for (i = 0; i < parallel_count; i++) | |||
165 | { | |||
166 | ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2]->data.u8); | |||
167 | ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2 + 1]->data.u8); | |||
168 | ccv_nnc_xpu_free(&compiled_data->xpu_alloc, max_normt[i]->data.u8); | |||
169 | } | |||
170 | for (i = 0; i < parallel_count; i++) | |||
171 | { | |||
172 | ccv_nnc_tensor_free(norm2[i * 2]); | |||
173 | ccv_nnc_tensor_free(norm2[i * 2 + 1]); | |||
174 | ccv_nnc_tensor_free(max_normt[i]); | |||
175 | } | |||
176 | } | |||
177 | ||||
178 | // MARK - Add-on Functions | |||
179 | ||||
180 | static int _ccv_cnnp_model_isnan(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context) | |||
181 | { | |||
182 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8); | |||
183 | ccv_nnc_tensor_t* const old_isnanr = outputs[1 + device_id * 2]; | |||
184 | ccv_nnc_tensor_t* const isnanr = outputs[1 + device_id * 2 + 1]; | |||
185 | ccv_nnc_cmd_t reduce_cmd = CMD_REDUCE_ISNAN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_ISNAN_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
186 | reduce_cmd.info.reduce.count = ccv_nnc_tensor_nd(inputs[0]->info.dim); | |||
187 | int i; | |||
188 | for (i = 0; i < cmd.info.reduce.count; i++) | |||
189 | reduce_cmd.info.reduce.axis[i] = i; | |||
190 | ccv_nnc_cmd_exec(reduce_cmd, hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(isnanr)(ccv_nnc_tensor_t* []){isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
191 | ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), hint, flags, TENSOR_LIST(old_isnanr, isnanr)(ccv_nnc_tensor_t* []){old_isnanr, isnanr}, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_isnanr)(ccv_nnc_tensor_t* []){old_isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
192 | return CCV_NNC_EXEC_SUCCESS; | |||
193 | } | |||
194 | ||||
195 | static ccv_nnc_cmd_vtab_t reduce_isnan_vtab = { | |||
196 | .exec = _ccv_cnnp_model_isnan | |||
197 | }; | |||
198 | ||||
199 | int ccv_cnnp_model_parameter_gradients_isnan(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, ccv_nnc_stream_context_t* const stream_context) | |||
200 | { | |||
201 | ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data; | |||
202 | assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c" , 202, __extension__ __PRETTY_FUNCTION__); })); | |||
203 | const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count ); typeof (1) _b = (1); (_a > _b) ? _a : _b; }); | |||
204 | ccv_nnc_tensor_t* isnanr[parallel_count * 2]; | |||
205 | const int stream_type = model->compiled_data->stream_type; | |||
206 | int i; | |||
207 | if (stream_type == CCV_STREAM_CONTEXT_GPU) | |||
208 | { | |||
209 | for (i = 0; i < parallel_count; i++) | |||
210 | { | |||
211 | ccv_nnc_tensor_param_t info = { | |||
212 | .type = CCV_TENSOR_GPU_MEMORY, | |||
213 | .format = CCV_TENSOR_FORMAT_NHWC, | |||
214 | .datatype = CCV_32S, | |||
215 | .dim = {1}, | |||
216 | }; | |||
217 | CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff ) << 8)); | |||
218 | isnanr[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0); | |||
219 | isnanr[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0); | |||
220 | } | |||
221 | } else { | |||
222 | for (i = 0; i < parallel_count; i++) | |||
223 | { | |||
224 | ccv_nnc_tensor_param_t info = { | |||
225 | .type = CCV_TENSOR_CPU_MEMORY, | |||
226 | .format = CCV_TENSOR_FORMAT_NHWC, | |||
227 | .datatype = CCV_32S, | |||
228 | .dim = {1}, | |||
229 | }; | |||
230 | isnanr[i * 2] = ccv_nnc_tensor_new(0, info, 0); | |||
231 | isnanr[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0); | |||
232 | } | |||
233 | } | |||
234 | // zero out old isnanr. | |||
235 | if (parallel_count > 1) | |||
236 | { | |||
237 | ccv_nnc_stream_context_t* streams[parallel_count]; | |||
238 | ccv_nnc_stream_signal_t* signal; | |||
239 | if (stream_context) | |||
240 | signal = ccv_nnc_stream_context_emit_signal_new(stream_context); | |||
241 | for (i = 0; i < parallel_count; i++) | |||
242 | { | |||
243 | const int stream_type = CCV_TENSOR_GET_MEMORY(isnanr[i * 2]->info.type)((isnanr[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU; | |||
244 | const int device_id = CCV_TENSOR_GET_DEVICE_ID(isnanr[i * 2]->info.type)(((isnanr[i * 2]->info.type) & 0xfff00) >> 8); | |||
245 | int type = stream_type; | |||
246 | CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff ) << 8)); | |||
247 | ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type); | |||
248 | // Wait signal to finish. | |||
249 | if (stream_context) | |||
250 | ccv_nnc_stream_context_wait_signal(stream_0, signal); | |||
251 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0); | |||
252 | if (stream_context) | |||
253 | { | |||
254 | ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0); | |||
255 | ccv_nnc_stream_context_wait_signal(stream_context, signal); | |||
256 | } | |||
257 | streams[i] = stream_0; | |||
258 | } | |||
259 | // If this should be blocking, blocking it. | |||
260 | if (!stream_context) | |||
261 | for (i = 0; i < parallel_count; i++) | |||
262 | if (streams[i]) | |||
263 | ccv_nnc_stream_context_wait(streams[i]); | |||
264 | } else | |||
265 | ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[0])(ccv_nnc_tensor_t* []){isnanr[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context); | |||
266 | // Gather isnanr. | |||
267 | ccv_nnc_cmd_t reduce_cmd = { | |||
268 | .cmd = CCV_NNC_CUSTOM_FORWARD, | |||
269 | .isa = &reduce_isnan_vtab, | |||
270 | }; | |||
271 | ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, isnanr, parallel_count * 2, stream_context); | |||
272 | for (i = 0; i < parallel_count; i++) | |||
273 | ccv_nnc_tensor_free(isnanr[i * 2 + 1]); | |||
274 | int retval = 0; | |||
275 | if (stream_type == CCV_TENSOR_GPU_MEMORY) | |||
276 | { | |||
277 | ccv_nnc_tensor_param_t info = { | |||
278 | .type = CCV_TENSOR_CPU_MEMORY, | |||
279 | .format = CCV_TENSOR_FORMAT_NHWC, | |||
280 | .datatype = CCV_32S, | |||
281 | .dim = {1}, | |||
282 | }; | |||
283 | ccv_nnc_tensor_t* checknan = ccv_nnc_tensor_new(0, info, 0); | |||
284 | for (i = 0; i < parallel_count; i++) | |||
285 | { | |||
286 | ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(checknan)(ccv_nnc_tensor_t* []){checknan}, (1 +1 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0); | |||
287 | if (checknan->data.i32[0] > 0) | |||
288 | { | |||
289 | retval = 1; | |||
290 | break; | |||
291 | } | |||
292 | } | |||
293 | ccv_nnc_tensor_free(checknan); | |||
294 | } else { | |||
295 | for (i = 0; i < parallel_count; i++) | |||
296 | if (isnanr[i * 2]->data.i32[0] > 0) | |||
297 | { | |||
298 | retval = 1; | |||
299 | break; | |||
300 | } | |||
301 | } | |||
302 | for (i = 0; i < parallel_count; i++) | |||
303 | ccv_nnc_tensor_free(isnanr[i * 2]); | |||
304 | return retval; | |||
305 | } | |||
306 | ||||
307 | // MARK - Core Layers | |||
308 | ||||
309 | static void _ccv_cnnp_sum_build(ccv_cnnp_model_t* const self, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
310 | { | |||
311 | PRINT(CCV_CLI_VERBOSE, "[cnnp_sum_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_sum_build] -\n"); fflush(stdout); } } while ( 0); | |||
312 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 312, __extension__ __PRETTY_FUNCTION__ ); })); | |||
313 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, ccv_nnc_tensor_symbol_params(graph, inputs[0]), 0); | |||
314 | ccv_nnc_graph_exec_symbol_new(graph, CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), inputs, input_size, outputs, output_size, 0); | |||
315 | } | |||
316 | ||||
317 | static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
318 | ||||
319 | static const ccv_cnnp_model_vtab_t ccv_cnnp_sum_isa = { | |||
320 | .build = _ccv_cnnp_sum_build, | |||
321 | .copy = _ccv_cnnp_sum_copy, | |||
322 | }; | |||
323 | ||||
324 | typedef struct { | |||
325 | ccv_cnnp_model_t super; | |||
326 | ccv_nnc_tensor_symbol_t output; | |||
327 | } ccv_cnnp_model_sum_t; | |||
328 | ||||
329 | ccv_cnnp_model_t* ccv_cnnp_sum(const char* const name) | |||
330 | { | |||
331 | ccv_cnnp_model_sum_t* const model_sum = (ccv_cnnp_model_sum_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sum_t)); | |||
332 | model_sum->super.isa = &ccv_cnnp_sum_isa; | |||
333 | model_sum->super.input_size = 0; | |||
334 | model_sum->super.outputs = &model_sum->output; | |||
335 | model_sum->super.output_size = 1; | |||
336 | ccv_cnnp_model_copy_name(&model_sum->super, name); | |||
337 | return (ccv_cnnp_model_t*)model_sum; | |||
338 | } | |||
339 | ||||
340 | static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
341 | { | |||
342 | return ccv_cnnp_sum(self->name); | |||
343 | } | |||
344 | ||||
345 | typedef struct { | |||
346 | ccv_cnnp_model_t super; | |||
347 | int axis; | |||
348 | ccv_nnc_tensor_symbol_t output; | |||
349 | } ccv_cnnp_model_concat_t; | |||
350 | ||||
351 | static void _ccv_cnnp_concat_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
352 | { | |||
353 | const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super; | |||
354 | PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] 1. -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_concat_build] 1. -\n"); fflush(stdout); } } while (0); | |||
| ||||
355 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 355, __extension__ __PRETTY_FUNCTION__ ); })); | |||
356 | ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
357 | int i, j; | |||
358 | if (output_params.dim[0] == 0) | |||
359 | for (i = 1; i < input_size; i++) | |||
360 | { | |||
361 | output_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
362 | if (output_params.dim[0] != 0) | |||
363 | break; | |||
364 | } | |||
365 | const int nd = ccv_nnc_tensor_nd(output_params.dim); | |||
366 | const int axis = self->axis; | |||
367 | assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if ( axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c" , 367, __extension__ __PRETTY_FUNCTION__); })); | |||
368 | output_params.dim[axis] = 0; | |||
369 | int input_is_contiguous = 1; | |||
370 | for (i = 0; i < input_size; i++) | |||
371 | { | |||
372 | const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
373 | const int input_nd = ccv_nnc_tensor_nd(input_params.dim); | |||
374 | if (input_nd == 0) | |||
375 | { | |||
376 | PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i) ; fflush(stdout); } } while (0); | |||
377 | input_is_contiguous = 0; | |||
378 | continue; | |||
379 | } | |||
380 | if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) | |||
381 | { | |||
382 | PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i, input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i, input_params.dim[0]); fflush(stdout); } } while (0); | |||
383 | int i; | |||
384 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++) | |||
385 | PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", input_params.dim[i]); fflush(stdout); } } while (0); | |||
386 | PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(")\n"); fflush(stdout); } } while (0); | |||
387 | } | |||
388 | assert(input_nd == nd)((void) sizeof ((input_nd == nd) ? 1 : 0), __extension__ ({ if (input_nd == nd) ; else __assert_fail ("input_nd == nd", "ccv_cnnp_model_addons.c" , 388, __extension__ __PRETTY_FUNCTION__); })); | |||
389 | for (j = 0; j < nd; j++) | |||
390 | if (j != axis) | |||
391 | { assert(input_params.dim[j] == output_params.dim[j])((void) sizeof ((input_params.dim[j] == output_params.dim[j]) ? 1 : 0), __extension__ ({ if (input_params.dim[j] == output_params .dim[j]) ; else __assert_fail ("input_params.dim[j] == output_params.dim[j]" , "ccv_cnnp_model_addons.c", 391, __extension__ __PRETTY_FUNCTION__ ); })); } | |||
392 | output_params.dim[axis] += input_params.dim[axis]; | |||
393 | } | |||
394 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
395 | int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
396 | int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
397 | ccv_nnc_tensor_get_stride(output_params.dim, stride); | |||
398 | if (input_is_contiguous
| |||
399 | { | |||
400 | ccv_nnc_tensor_symbol_t aliases[input_size]; | |||
| ||||
401 | for (i = 0; i < input_size; i++) | |||
402 | { | |||
403 | const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
404 | aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0); | |||
405 | ofs[axis] += input_params.dim[axis]; | |||
406 | } | |||
407 | // Format transform is more flexible. | |||
408 | ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), inputs, input_size, aliases, input_size, "concat"); | |||
409 | } else { | |||
410 | ccv_nnc_tensor_symbol_t aliases[input_size]; | |||
411 | for (i = 0; i < input_size; i++) | |||
412 | { | |||
413 | const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
414 | if (input_params.dim[0] == 0) | |||
415 | { | |||
416 | // Create a new alias anyway, but not going to use it, in this way, the alias count will match during absorb. | |||
417 | aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0); | |||
418 | continue; | |||
419 | } | |||
420 | aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0); | |||
421 | ofs[axis] += input_params.dim[axis]; | |||
422 | } | |||
423 | // Format transform is more flexible. | |||
424 | ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), inputs, input_size, aliases, input_size, "concat"); | |||
425 | } | |||
426 | } | |||
427 | ||||
428 | static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
429 | ||||
430 | static const ccv_cnnp_model_vtab_t ccv_cnnp_concat_isa = { | |||
431 | .build = _ccv_cnnp_concat_build, | |||
432 | .copy = _ccv_cnnp_concat_copy, | |||
433 | }; | |||
434 | ||||
435 | ccv_cnnp_model_t* ccv_cnnp_concat(const int axis, const char* const name) | |||
436 | { | |||
437 | ccv_cnnp_model_concat_t* const model_concat = (ccv_cnnp_model_concat_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_concat_t)); | |||
438 | model_concat->super.isa = &ccv_cnnp_concat_isa; | |||
439 | model_concat->super.input_size = 0; | |||
440 | model_concat->super.outputs = &model_concat->output; | |||
441 | model_concat->super.output_size = 1; | |||
442 | model_concat->axis = axis; | |||
443 | ccv_cnnp_model_copy_name(&model_concat->super, name); | |||
444 | return (ccv_cnnp_model_t*)model_concat; | |||
445 | } | |||
446 | ||||
447 | static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
448 | { | |||
449 | const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super; | |||
450 | return ccv_cnnp_concat(self->axis, self->super.name); | |||
451 | } | |||
452 | ||||
453 | typedef struct { | |||
454 | ccv_cnnp_model_t super; | |||
455 | int axis; | |||
456 | ccv_nnc_tensor_symbol_t outputs[1]; | |||
457 | } ccv_cnnp_model_chunk_t; | |||
458 | ||||
459 | static void _ccv_cnnp_chunk_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
460 | { | |||
461 | const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super; | |||
462 | PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 1. axis: %d\n", self->axis)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_chunk_build] 1. axis: %d\n", self->axis); fflush(stdout); } } while (0); | |||
463 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 463, __extension__ __PRETTY_FUNCTION__); })); | |||
464 | const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
465 | if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) | |||
466 | { | |||
467 | PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 2. input: (%d", input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_chunk_build] 2. input: (%d", input_params.dim [0]); fflush(stdout); } } while (0); | |||
468 | int i; | |||
469 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++) | |||
470 | PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", input_params.dim[i]); fflush(stdout); } } while (0); | |||
471 | PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(")\n"); fflush(stdout); } } while (0); | |||
472 | } | |||
473 | ccv_nnc_tensor_param_t output_params = input_params; | |||
474 | int i; | |||
475 | const int nd = ccv_nnc_tensor_nd(output_params.dim); | |||
476 | const int axis = self->axis; | |||
477 | assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if ( axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c" , 477, __extension__ __PRETTY_FUNCTION__); })); | |||
478 | const int n = self->super.output_size; | |||
479 | assert(n == output_size)((void) sizeof ((n == output_size) ? 1 : 0), __extension__ ({ if (n == output_size) ; else __assert_fail ("n == output_size" , "ccv_cnnp_model_addons.c", 479, __extension__ __PRETTY_FUNCTION__ ); })); | |||
480 | assert(output_params.dim[axis] % n == 0)((void) sizeof ((output_params.dim[axis] % n == 0) ? 1 : 0), __extension__ ({ if (output_params.dim[axis] % n == 0) ; else __assert_fail ("output_params.dim[axis] % n == 0", "ccv_cnnp_model_addons.c" , 480, __extension__ __PRETTY_FUNCTION__); })); | |||
481 | output_params.dim[axis] = output_params.dim[axis] / n; | |||
482 | int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
483 | int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
484 | ccv_nnc_tensor_get_stride(input_params.dim, stride); | |||
485 | ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]); | |||
486 | if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward. | |||
487 | { | |||
488 | for (i = 0; i < output_size; i++) | |||
489 | { | |||
490 | outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0); | |||
491 | ofs[axis] += output_params.dim[axis]; | |||
492 | } | |||
493 | } else { | |||
494 | // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly. | |||
495 | // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong. | |||
496 | int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
497 | ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride); | |||
498 | // We identify permute by checking if the stride is not in descending order. | |||
499 | // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly. | |||
500 | int i, no_permute = 1; | |||
501 | for (i = 1; no_permute && i < nd; i++) | |||
502 | if (old_stride[i - 1] < old_stride[i]) | |||
503 | no_permute = 0; | |||
504 | if (no_permute) | |||
505 | { // Just straightforward reshape if there is no no permute. | |||
506 | for (i = 0; i < output_size; i++) | |||
507 | { | |||
508 | outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, old_stride, output_params, 0); | |||
509 | ofs[axis] += output_params.dim[axis]; | |||
510 | } | |||
511 | } else { | |||
512 | // Otherwise, we first do format transform to plain tensor and then do reshape. | |||
513 | ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, input_params, 0); | |||
514 | ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape"); | |||
515 | for (i = 0; i < output_size; i++) | |||
516 | { | |||
517 | outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, ofs, stride, output_params, 0); | |||
518 | ofs[axis] += output_params.dim[axis]; | |||
519 | } | |||
520 | } | |||
521 | } | |||
522 | } | |||
523 | ||||
524 | static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
525 | ||||
526 | static const ccv_cnnp_model_vtab_t ccv_cnnp_chunk_isa = { | |||
527 | .build = _ccv_cnnp_chunk_build, | |||
528 | .copy = _ccv_cnnp_chunk_copy, | |||
529 | }; | |||
530 | ||||
531 | ccv_cnnp_model_t* ccv_cnnp_chunk(const int n, const int axis, const char* const name) | |||
532 | { | |||
533 | assert(n >= 1)((void) sizeof ((n >= 1) ? 1 : 0), __extension__ ({ if (n >= 1) ; else __assert_fail ("n >= 1", "ccv_cnnp_model_addons.c" , 533, __extension__ __PRETTY_FUNCTION__); })); | |||
534 | ccv_cnnp_model_chunk_t* const model_chunk = (ccv_cnnp_model_chunk_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_chunk_t) + sizeof(ccv_nnc_tensor_symbol_t) * (n - 1)); | |||
535 | model_chunk->super.isa = &ccv_cnnp_chunk_isa; | |||
536 | model_chunk->super.input_size = 1; | |||
537 | model_chunk->super.outputs = model_chunk->outputs; | |||
538 | model_chunk->super.output_size = n; | |||
539 | model_chunk->axis = axis; | |||
540 | ccv_cnnp_model_copy_name(&model_chunk->super, name); | |||
541 | return (ccv_cnnp_model_t*)model_chunk; | |||
542 | } | |||
543 | ||||
544 | static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
545 | { | |||
546 | const ccv_cnnp_model_chunk_t* const self = (const ccv_cnnp_model_chunk_t*)super; | |||
547 | return ccv_cnnp_chunk(self->super.output_size, self->axis, self->super.name); | |||
548 | } | |||
549 | ||||
550 | typedef struct { | |||
551 | ccv_cnnp_model_t super; | |||
552 | ccv_nnc_tensor_symbol_t output; | |||
553 | int format; | |||
554 | int dim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
555 | int ofs[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
556 | int stride[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
557 | } ccv_cnnp_model_reshape_t; | |||
558 | ||||
559 | static void _ccv_cnnp_reshape_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
560 | { | |||
561 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 561, __extension__ __PRETTY_FUNCTION__); })); | |||
562 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 562, __extension__ __PRETTY_FUNCTION__ ); })); | |||
563 | ccv_cnnp_model_reshape_t* const self = (ccv_cnnp_model_reshape_t*)super; | |||
564 | if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) | |||
565 | { | |||
566 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 1. dim: (%d", self->dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reshape_build] 1. dim: (%d", self->dim[0] ); fflush(stdout); } } while (0); | |||
567 | int i; | |||
568 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && self->dim[i] > 0; i++) | |||
569 | PRINT(CCV_CLI_VERBOSE, ", %d", self->dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", self->dim[i]); fflush(stdout); } } while (0); | |||
570 | const int count = i; | |||
571 | PRINT(CCV_CLI_VERBOSE, "), ofs: (%d", self->ofs[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("), ofs: (%d", self->ofs[0]); fflush(stdout); } } while (0); | |||
572 | for (i = 1; i < count; i++) | |||
573 | PRINT(CCV_CLI_VERBOSE, ", %d", self->ofs[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", self->ofs[i]); fflush(stdout); } } while (0); | |||
574 | PRINT(CCV_CLI_VERBOSE, "), stride: (%d", self->stride[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("), stride: (%d", self->stride[0]); fflush(stdout ); } } while (0); | |||
575 | for (i = 1; i < count; i++) | |||
576 | PRINT(CCV_CLI_VERBOSE, ", %d", self->stride[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", self->stride[i]); fflush(stdout); } } while (0); | |||
577 | PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(")\n"); fflush(stdout); } } while (0); | |||
578 | } | |||
579 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
580 | if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) | |||
581 | { | |||
582 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 2. input: (%d", params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reshape_build] 2. input: (%d", params.dim[0] ); fflush(stdout); } } while (0); | |||
583 | int i; | |||
584 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && params.dim[i] > 0; i++) | |||
585 | PRINT(CCV_CLI_VERBOSE, ", %d", params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(", %d", params.dim[i]); fflush(stdout); } } while ( 0); | |||
586 | PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(")\n"); fflush(stdout); } } while (0); | |||
587 | } | |||
588 | if (self->format > 0) | |||
589 | params.format = self->format; | |||
590 | assert(ccv_nnc_dimension_count(self->dim) <= ccv_nnc_tensor_count(params))((void) sizeof ((ccv_nnc_dimension_count(self->dim) <= ccv_nnc_tensor_count (params)) ? 1 : 0), __extension__ ({ if (ccv_nnc_dimension_count (self->dim) <= ccv_nnc_tensor_count(params)) ; else __assert_fail ("ccv_nnc_dimension_count(self->dim) <= ccv_nnc_tensor_count(params)" , "ccv_cnnp_model_addons.c", 590, __extension__ __PRETTY_FUNCTION__ ); })); | |||
591 | ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]); | |||
592 | int stride_from_dim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
593 | if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward. | |||
594 | { | |||
595 | memcpy(params.dim, self->dim, sizeof(params.dim)); | |||
596 | int* stride; | |||
597 | if (self->stride[0] == 0) | |||
598 | { | |||
599 | ccv_nnc_tensor_get_stride(self->dim, stride_from_dim); | |||
600 | stride = stride_from_dim; | |||
601 | } else | |||
602 | stride = self->stride; | |||
603 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0); | |||
604 | } else { | |||
605 | // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly. | |||
606 | // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong. | |||
607 | int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
608 | ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride); | |||
609 | // We identify permute by checking if the stride is not in descending order. | |||
610 | // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly. | |||
611 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
612 | const int new_nd = ccv_nnc_tensor_nd(self->dim); | |||
613 | int i, no_permute = 1; | |||
614 | // If the new dim has different nd, or we actually have a stride, we need to check if it is no permute or not. | |||
615 | if (new_nd != nd || (self->stride[0] != 0 && memcmp(self->stride, old_stride, sizeof(self->stride)) != 0)) | |||
616 | for (i = 1; no_permute && i < nd; i++) | |||
617 | if (old_stride[i - 1] < old_stride[i]) | |||
618 | no_permute = 0; | |||
619 | if (no_permute) | |||
620 | { // Just straightforward reshape if there is no no permute. | |||
621 | memcpy(params.dim, self->dim, sizeof(params.dim)); | |||
622 | int* stride; | |||
623 | if (self->stride[0] == 0) | |||
624 | { | |||
625 | if (new_nd != nd) // Cannot use old stride. | |||
626 | { | |||
627 | ccv_nnc_tensor_get_stride(self->dim, stride_from_dim); | |||
628 | stride = stride_from_dim; | |||
629 | } else | |||
630 | stride = old_stride; | |||
631 | } else | |||
632 | stride = self->stride; | |||
633 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0); | |||
634 | } else { | |||
635 | // Otherwise, we first do format transform to plain tensor and then do reshape. | |||
636 | ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
637 | ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape"); | |||
638 | memcpy(params.dim, self->dim, sizeof(params.dim)); | |||
639 | int* stride; | |||
640 | if (self->stride[0] == 0) | |||
641 | { | |||
642 | ccv_nnc_tensor_get_stride(self->dim, stride_from_dim); | |||
643 | stride = stride_from_dim; | |||
644 | } else | |||
645 | stride = self->stride; | |||
646 | // And then we create alias against the permuted one. | |||
647 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, self->ofs, stride, params, 0); | |||
648 | } | |||
649 | } | |||
650 | } | |||
651 | ||||
652 | static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
653 | ||||
654 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reshape_isa = { | |||
655 | .build = _ccv_cnnp_reshape_build, | |||
656 | .copy = _ccv_cnnp_reshape_copy, | |||
657 | }; | |||
658 | ||||
659 | ccv_cnnp_model_t* ccv_cnnp_reshape(const int format, const int dim[CCV_NNC_MAX_DIM_ALLOC(12)], const int ofs[CCV_NNC_MAX_DIM_ALLOC(12)], const int stride[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name) | |||
660 | { | |||
661 | ccv_cnnp_model_reshape_t* const model_reshape = (ccv_cnnp_model_reshape_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reshape_t)); | |||
662 | model_reshape->super.isa = &ccv_cnnp_reshape_isa; | |||
663 | model_reshape->super.input_size = 1; | |||
664 | model_reshape->super.outputs = &model_reshape->output; | |||
665 | model_reshape->super.output_size = 1; | |||
666 | ccv_cnnp_model_copy_name(&model_reshape->super, name); | |||
667 | model_reshape->format = format; | |||
668 | memcpy(model_reshape->dim, dim, sizeof(model_reshape->dim)); | |||
669 | memcpy(model_reshape->ofs, ofs, sizeof(model_reshape->ofs)); | |||
670 | if (stride[0] != 0) | |||
671 | memcpy(model_reshape->stride, stride, sizeof(model_reshape->stride)); | |||
672 | return (ccv_cnnp_model_t*)model_reshape; | |||
673 | } | |||
674 | ||||
675 | static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
676 | { | |||
677 | const ccv_cnnp_model_reshape_t* const self = (const ccv_cnnp_model_reshape_t*)super; | |||
678 | return ccv_cnnp_reshape(self->format, self->dim, self->ofs, self->stride, self->super.name); | |||
679 | } | |||
680 | ||||
681 | typedef struct { | |||
682 | ccv_cnnp_model_t super; | |||
683 | ccv_nnc_tensor_symbol_t output; | |||
684 | int type; | |||
685 | int begin[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
686 | int end[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
687 | } ccv_cnnp_model_pad_t; | |||
688 | ||||
689 | static void _ccv_cnnp_pad_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
690 | { | |||
691 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 691, __extension__ __PRETTY_FUNCTION__); })); | |||
692 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 692, __extension__ __PRETTY_FUNCTION__ ); })); | |||
693 | ccv_cnnp_model_pad_t* const self = (ccv_cnnp_model_pad_t*)super; | |||
694 | PRINT(CCV_CLI_VERBOSE, "[cnnp_pad_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_pad_build] -\n"); fflush(stdout); } } while ( 0); | |||
695 | const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
696 | const int nd = ccv_nnc_tensor_nd(input_params.dim); | |||
697 | ccv_nnc_tensor_param_t params = input_params; | |||
698 | int i; | |||
699 | for (i = 0 ; i < nd; i++) | |||
700 | params.dim[i] += self->begin[i] + self->end[i]; | |||
701 | const ccv_nnc_tensor_symbol_t padded = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
702 | ccv_nnc_cmd_t pad = CMD_PAD_FORWARD(self->type, (), ())ccv_nnc_cmd(CCV_NNC_PAD_FORWARD, 0, ((ccv_nnc_cmd_param_t){.size ={.dim={}},.pad={.type=self->type,.end={}}}), 0); | |||
703 | memcpy(pad.info.size.dim, self->begin, sizeof(pad.info.size.dim)); | |||
704 | memcpy(pad.info.pad.end, self->end, sizeof(pad.info.pad.end)); | |||
705 | ccv_nnc_graph_exec_symbol_new(graph, pad, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(padded)(const ccv_nnc_tensor_symbol_t []){padded}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "pad"); | |||
706 | outputs[0] = padded; | |||
707 | } | |||
708 | ||||
709 | static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
710 | ||||
711 | static const ccv_cnnp_model_vtab_t ccv_cnnp_pad_isa = { | |||
712 | .build = _ccv_cnnp_pad_build, | |||
713 | .copy = _ccv_cnnp_pad_copy, | |||
714 | }; | |||
715 | ||||
716 | ccv_cnnp_model_t* ccv_cnnp_pad(const int type, const int begin[CCV_NNC_MAX_DIM_ALLOC(12)], const int end[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name) | |||
717 | { | |||
718 | ccv_cnnp_model_pad_t* const model_pad = (ccv_cnnp_model_pad_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pad_t)); | |||
719 | model_pad->super.isa = &ccv_cnnp_pad_isa; | |||
720 | model_pad->super.input_size = 1; | |||
721 | model_pad->super.outputs = &model_pad->output; | |||
722 | model_pad->super.output_size = 1; | |||
723 | ccv_cnnp_model_copy_name(&model_pad->super, name); | |||
724 | model_pad->type = type; | |||
725 | memcpy(model_pad->begin, begin, sizeof(model_pad->begin)); | |||
726 | memcpy(model_pad->end, end, sizeof(model_pad->end)); | |||
727 | return (ccv_cnnp_model_t*)model_pad; | |||
728 | } | |||
729 | ||||
730 | static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
731 | { | |||
732 | const ccv_cnnp_model_pad_t* const self = (const ccv_cnnp_model_pad_t*)super; | |||
733 | return ccv_cnnp_pad(self->type, self->begin, self->end, self->super.name); | |||
734 | } | |||
735 | ||||
736 | typedef struct { | |||
737 | ccv_cnnp_model_t super; | |||
738 | ccv_nnc_tensor_symbol_t output; | |||
739 | } ccv_cnnp_model_identity_t; | |||
740 | ||||
741 | static void _ccv_cnnp_identity_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
742 | { | |||
743 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 743, __extension__ __PRETTY_FUNCTION__); })); | |||
744 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 744, __extension__ __PRETTY_FUNCTION__ ); })); | |||
745 | PRINT(CCV_CLI_VERBOSE, "[cnnp_identity_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_identity_build] -\n"); fflush(stdout); } } while (0); | |||
746 | outputs[0] = inputs[0]; | |||
747 | } | |||
748 | ||||
749 | static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
750 | ||||
751 | static const ccv_cnnp_model_vtab_t ccv_cnnp_identity_isa = { | |||
752 | .build = _ccv_cnnp_identity_build, | |||
753 | .copy = _ccv_cnnp_identity_copy, | |||
754 | }; | |||
755 | ||||
756 | ccv_cnnp_model_t* ccv_cnnp_identity(const char* const name) | |||
757 | { | |||
758 | ccv_cnnp_model_identity_t* const model_identity = (ccv_cnnp_model_identity_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_identity_t)); | |||
759 | model_identity->super.isa = &ccv_cnnp_identity_isa; | |||
760 | model_identity->super.input_size = 1; | |||
761 | model_identity->super.outputs = &model_identity->output; | |||
762 | model_identity->super.output_size = 1; | |||
763 | ccv_cnnp_model_copy_name(&model_identity->super, name); | |||
764 | return (ccv_cnnp_model_t*)model_identity; | |||
765 | } | |||
766 | ||||
767 | static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
768 | { | |||
769 | const ccv_cnnp_model_identity_t* const self = (const ccv_cnnp_model_identity_t*)super; | |||
770 | return ccv_cnnp_identity(self->super.name); | |||
771 | } | |||
772 | ||||
773 | typedef struct { | |||
774 | ccv_cnnp_model_t super; | |||
775 | ccv_nnc_tensor_symbol_t output; | |||
776 | int index[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
777 | } ccv_cnnp_model_permute_t; | |||
778 | ||||
779 | static void _ccv_cnnp_permute_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
780 | { | |||
781 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 781, __extension__ __PRETTY_FUNCTION__); })); | |||
782 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 782, __extension__ __PRETTY_FUNCTION__ ); })); | |||
783 | ccv_cnnp_model_permute_t* const self = (ccv_cnnp_model_permute_t*)super; | |||
784 | PRINT(CCV_CLI_VERBOSE, "[cnnp_permute_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_permute_build] -\n"); fflush(stdout); } } while (0); | |||
785 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
786 | ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]); | |||
787 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
788 | int input_dim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
789 | memcpy(input_dim, params.dim, sizeof(params.dim)); | |||
790 | int input_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
791 | int output_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
792 | if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If it is not an alias. Find stride and permute. | |||
793 | { | |||
794 | ccv_nnc_tensor_get_stride(input_dim, input_stride); | |||
795 | int i; | |||
796 | for (i = 0; i < nd; i++) | |||
797 | { | |||
798 | const int idx = self->index[i]; | |||
799 | assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0) , __extension__ ({ if (idx >= 0 && idx < nd) ; else __assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c" , 799, __extension__ __PRETTY_FUNCTION__); })); | |||
800 | params.dim[i] = input_dim[idx]; | |||
801 | output_stride[i] = input_stride[idx]; | |||
802 | } | |||
803 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ccv_nnc_no_ofs, output_stride, params, 0); | |||
804 | } else { | |||
805 | // if it is an alias, we can get the stride from it and use that. | |||
806 | int input_ofs[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
807 | ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], input_ofs, input_stride); | |||
808 | assert(input_stride[0] != 0)((void) sizeof ((input_stride[0] != 0) ? 1 : 0), __extension__ ({ if (input_stride[0] != 0) ; else __assert_fail ("input_stride[0] != 0" , "ccv_cnnp_model_addons.c", 808, __extension__ __PRETTY_FUNCTION__ ); })); | |||
809 | int output_ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
810 | int i; | |||
811 | for (i = 0; i < nd; i++) | |||
812 | { | |||
813 | const int idx = self->index[i]; | |||
814 | assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0) , __extension__ ({ if (idx >= 0 && idx < nd) ; else __assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c" , 814, __extension__ __PRETTY_FUNCTION__); })); | |||
815 | params.dim[i] = input_dim[idx]; | |||
816 | output_stride[i] = input_stride[idx]; | |||
817 | output_ofs[i] = input_ofs[idx]; | |||
818 | } | |||
819 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], output_ofs, output_stride, params, 0); | |||
820 | } | |||
821 | } | |||
822 | ||||
823 | static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
824 | ||||
825 | static const ccv_cnnp_model_vtab_t ccv_cnnp_permute_isa = { | |||
826 | .build = _ccv_cnnp_permute_build, | |||
827 | .copy = _ccv_cnnp_permute_copy, | |||
828 | }; | |||
829 | ||||
830 | ccv_cnnp_model_t* ccv_cnnp_permute(const int index[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name) | |||
831 | { | |||
832 | ccv_cnnp_model_permute_t* const model_permute = (ccv_cnnp_model_permute_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_permute_t)); | |||
833 | model_permute->super.isa = &ccv_cnnp_permute_isa; | |||
834 | model_permute->super.input_size = 1; | |||
835 | model_permute->super.outputs = &model_permute->output; | |||
836 | model_permute->super.output_size = 1; | |||
837 | ccv_cnnp_model_copy_name(&model_permute->super, name); | |||
838 | memcpy(model_permute->index, index, sizeof(model_permute->index)); | |||
839 | return (ccv_cnnp_model_t*)model_permute; | |||
840 | } | |||
841 | ||||
842 | static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
843 | { | |||
844 | const ccv_cnnp_model_permute_t* const self = (const ccv_cnnp_model_permute_t*)super; | |||
845 | return ccv_cnnp_permute(self->index, self->super.name); | |||
846 | } | |||
847 | ||||
848 | typedef struct { | |||
849 | ccv_cnnp_model_t super; | |||
850 | int index; | |||
851 | ccv_nnc_tensor_symbol_t output; | |||
852 | } ccv_cnnp_model_extract_t; | |||
853 | ||||
854 | static void _ccv_cnnp_extract_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
855 | { | |||
856 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 856, __extension__ __PRETTY_FUNCTION__ ); })); | |||
857 | ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super; | |||
858 | PRINT(CCV_CLI_VERBOSE, "[cnnp_extract_build] index: %d\n", self->index)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_extract_build] index: %d\n", self->index) ; fflush(stdout); } } while (0); | |||
859 | outputs[0] = inputs[self->index]; | |||
860 | } | |||
861 | ||||
862 | static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
863 | ||||
864 | static const ccv_cnnp_model_vtab_t ccv_cnnp_extract_isa = { | |||
865 | .build = _ccv_cnnp_extract_build, | |||
866 | .copy = _ccv_cnnp_extract_copy, | |||
867 | }; | |||
868 | ||||
869 | ccv_cnnp_model_t* ccv_cnnp_extract(const int index, const char* const name) | |||
870 | { | |||
871 | ccv_cnnp_model_extract_t* const model_extract = (ccv_cnnp_model_extract_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_extract_t)); | |||
872 | model_extract->index = index; | |||
873 | model_extract->super.isa = &ccv_cnnp_extract_isa; | |||
874 | model_extract->super.input_size = 0; | |||
875 | model_extract->super.outputs = &model_extract->output; | |||
876 | model_extract->super.output_size = 1; | |||
877 | ccv_cnnp_model_copy_name(&model_extract->super, name); | |||
878 | return (ccv_cnnp_model_t*)model_extract; | |||
879 | } | |||
880 | ||||
881 | static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
882 | { | |||
883 | ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super; | |||
884 | return ccv_cnnp_extract(self->index, self->super.name); | |||
885 | } | |||
886 | ||||
887 | typedef struct { | |||
888 | ccv_cnnp_model_t super; | |||
889 | ccv_nnc_tensor_symbol_t output; | |||
890 | } ccv_cnnp_model_flatten_t; | |||
891 | ||||
892 | static void _ccv_cnnp_flatten_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
893 | { | |||
894 | PRINT(CCV_CLI_VERBOSE, "[cnnp_flatten_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_flatten_build] -\n"); fflush(stdout); } } while (0); | |||
895 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 895, __extension__ __PRETTY_FUNCTION__); })); | |||
896 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 896, __extension__ __PRETTY_FUNCTION__ ); })); | |||
897 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
898 | ccv_nnc_tensor_param_t output_params = params; | |||
899 | memset(output_params.dim, 0, sizeof(output_params.dim)); | |||
900 | output_params.dim[0] = ccv_nnc_tensor_get_n(params); | |||
901 | assert(output_params.dim[0] > 0)((void) sizeof ((output_params.dim[0] > 0) ? 1 : 0), __extension__ ({ if (output_params.dim[0] > 0) ; else __assert_fail ("output_params.dim[0] > 0" , "ccv_cnnp_model_addons.c", 901, __extension__ __PRETTY_FUNCTION__ ); })); | |||
902 | output_params.dim[1] = ccv_nnc_tensor_count(params) / output_params.dim[0]; | |||
903 | int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
904 | ccv_nnc_tensor_get_stride(output_params.dim, stride); | |||
905 | outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], DIM_ALLOC()(int [(12)]){}, stride, output_params, 0); | |||
906 | } | |||
907 | ||||
908 | static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
909 | ||||
910 | static const ccv_cnnp_model_vtab_t ccv_cnnp_flatten_isa = { | |||
911 | .build = _ccv_cnnp_flatten_build, | |||
912 | .copy = _ccv_cnnp_flatten_copy, | |||
913 | }; | |||
914 | ||||
915 | ccv_cnnp_model_t* ccv_cnnp_flatten(const char* const name) | |||
916 | { | |||
917 | ccv_cnnp_model_flatten_t* const model_flatten = (ccv_cnnp_model_flatten_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_flatten_t)); | |||
918 | model_flatten->super.isa = &ccv_cnnp_flatten_isa; | |||
919 | model_flatten->super.input_size = 1; | |||
920 | model_flatten->super.outputs = &model_flatten->output; | |||
921 | model_flatten->super.output_size = 1; | |||
922 | ccv_cnnp_model_copy_name(&model_flatten->super, name); | |||
923 | return (ccv_cnnp_model_t*)model_flatten; | |||
924 | } | |||
925 | ||||
926 | static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
927 | { | |||
928 | return ccv_cnnp_flatten(self->name); | |||
929 | } | |||
930 | ||||
931 | // MARK - Batch Norm Layer | |||
932 | ||||
933 | typedef struct { | |||
934 | ccv_cnnp_model_t super; | |||
935 | ccv_nnc_tensor_symbol_t output; | |||
936 | ccv_nnc_tensor_symbol_t bias; | |||
937 | ccv_nnc_tensor_symbol_t scale; | |||
938 | ccv_nnc_graph_exec_symbol_t batch_norm; | |||
939 | ccv_nnc_cmd_param_t params; | |||
940 | ccv_array_t* zero_inits; | |||
941 | ccv_array_t* retainables; | |||
942 | } ccv_cnnp_model_batch_norm_t; | |||
943 | ||||
944 | static void _ccv_cnnp_batch_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
945 | { | |||
946 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 946, __extension__ __PRETTY_FUNCTION__); })); | |||
947 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 947, __extension__ __PRETTY_FUNCTION__ ); })); | |||
948 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
949 | PRINT(CCV_CLI_VERBOSE, "[cnnp_batch_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_batch_norm_build] -\n"); fflush(stdout); } } while (0); | |||
950 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
951 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
952 | ccv_nnc_tensor_param_t bias_params = params; | |||
953 | memset(bias_params.dim, 0, sizeof(bias_params.dim)); | |||
954 | // If the accuracy is not enough, bump it to 32-bit floating point. | |||
955 | if (bias_params.datatype != CCV_32F && bias_params.datatype != CCV_64F) | |||
956 | bias_params.datatype = CCV_32F; | |||
957 | bias_params.dim[0] = nd > 1 ? ccv_nnc_tensor_get_c(params) : params.dim[0]; | |||
958 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
959 | // Both scale and bias are shared between if this model is reused. | |||
960 | if (!self->scale.graph) | |||
961 | self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale"); | |||
962 | if (!self->bias.graph) | |||
963 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
964 | const ccv_nnc_tensor_symbol_t mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "mean"); | |||
965 | const ccv_nnc_tensor_symbol_t var = ccv_nnc_tensor_symbol_new(graph, bias_params, "var"); | |||
966 | // Otherwise, notice mean, var, saved_mean, saved_inv_std are not reused. | |||
967 | if (!self->zero_inits) | |||
968 | self->zero_inits = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
969 | ccv_array_push(self->zero_inits, &mean); | |||
970 | ccv_array_push(self->zero_inits, &var); | |||
971 | const ccv_nnc_tensor_symbol_t out_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_mean"); | |||
972 | const ccv_nnc_tensor_symbol_t out_var = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_var"); | |||
973 | if (!self->retainables) | |||
974 | self->retainables = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
975 | ccv_array_push(self->retainables, &out_mean); | |||
976 | ccv_array_push(self->retainables, &out_var); | |||
977 | const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_mean"); | |||
978 | const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_inv_std"); | |||
979 | const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim)); | |||
980 | ccv_nnc_cmd_param_t batch_norm = self->params; | |||
981 | batch_norm.bnorm.count = hw >= 0 ? CCV_NNC_MAX_DIM(2) + 1 : 1; | |||
982 | int i; | |||
983 | batch_norm.bnorm.axis[0] = (params.format == CCV_TENSOR_FORMAT_CHWN) ? 3 : 0; | |||
984 | if (hw >= 0) | |||
985 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
986 | batch_norm.bnorm.axis[i + 1] = i + hw; | |||
987 | self->params = batch_norm; | |||
988 | self->batch_norm = ccv_nnc_graph_exec_symbol_new(graph, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, batch_norm, 0), TENSOR_SYMBOL_LIST(inputs[0], self->scale, self->bias, mean, var)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->scale, self->bias, mean, var}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, out_mean, out_var, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, out_mean, out_var, saved_mean, saved_inv_std}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "batch_norm"); | |||
989 | outputs[0] = output; | |||
990 | } | |||
991 | ||||
992 | static void _ccv_cnnp_batch_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
993 | { | |||
994 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
995 | if (self->scale.graph) | |||
996 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(0, 1)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={0, 1}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale); | |||
997 | if (self->bias.graph) | |||
998 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
999 | int i; | |||
1000 | if (self->zero_inits) | |||
1001 | for (i = 0; i < self->zero_inits->rnum; i++) | |||
1002 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->zero_inits, i)((void*)(((char*)((self->zero_inits)->data)) + (size_t) (self->zero_inits)->rsize * (size_t)(i)))); | |||
1003 | } | |||
1004 | ||||
1005 | static void _ccv_cnnp_batch_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
1006 | { | |||
1007 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
1008 | if (self->scale.graph) | |||
1009 | add_to_array(parameters, self->scale, is_trainable); | |||
1010 | if (self->bias.graph) | |||
1011 | add_to_array(parameters, self->bias, is_trainable); | |||
1012 | } | |||
1013 | ||||
1014 | static void _ccv_cnnp_batch_norm_add_to_output(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const outputs) | |||
1015 | { | |||
1016 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
1017 | int i; | |||
1018 | if (self->retainables) | |||
1019 | for (i = 0; i < self->retainables->rnum; i++) | |||
1020 | { | |||
1021 | const ccv_nnc_tensor_symbol_t symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->retainables, i)((void*)(((char*)((self->retainables)->data)) + (size_t )(self->retainables)->rsize * (size_t)(i))); | |||
1022 | add_to_array(outputs, symbol, 0); | |||
1023 | } | |||
1024 | } | |||
1025 | ||||
1026 | static void _ccv_cnnp_batch_norm_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context) | |||
1027 | { | |||
1028 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
1029 | if (self->batch_norm.graph) | |||
1030 | { | |||
1031 | self->params.bnorm.is_test = is_test; | |||
1032 | updater(context, self->batch_norm, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint); | |||
1033 | } | |||
1034 | } | |||
1035 | ||||
1036 | static void _ccv_cnnp_batch_norm_deinit(ccv_cnnp_model_t* const super) | |||
1037 | { | |||
1038 | ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super; | |||
1039 | if (self->zero_inits) | |||
1040 | ccv_array_free(self->zero_inits); | |||
1041 | if (self->retainables) | |||
1042 | ccv_array_free(self->retainables); | |||
1043 | } | |||
1044 | ||||
1045 | static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1046 | ||||
1047 | static const ccv_cnnp_model_vtab_t ccv_cnnp_batch_norm_isa = { | |||
1048 | .build = _ccv_cnnp_batch_norm_build, | |||
1049 | .init_states = _ccv_cnnp_batch_norm_init_states, | |||
1050 | .add_to_parameter = _ccv_cnnp_batch_norm_add_to_parameter, | |||
1051 | .add_to_output = _ccv_cnnp_batch_norm_add_to_output, | |||
1052 | .copy = _ccv_cnnp_batch_norm_copy, | |||
1053 | .set_is_test = _ccv_cnnp_batch_norm_set_is_test, | |||
1054 | .deinit = _ccv_cnnp_batch_norm_deinit, | |||
1055 | }; | |||
1056 | ||||
1057 | ccv_cnnp_model_t* ccv_cnnp_batch_norm(const float momentum, const float epsilon, const int is_trainable, const char* const name) | |||
1058 | { | |||
1059 | ccv_cnnp_model_batch_norm_t* const model_batch_norm = (ccv_cnnp_model_batch_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_batch_norm_t)); | |||
1060 | model_batch_norm->super.isa = &ccv_cnnp_batch_norm_isa; | |||
1061 | model_batch_norm->super.input_size = 1; | |||
1062 | model_batch_norm->super.outputs = &model_batch_norm->output; | |||
1063 | model_batch_norm->super.output_size = 1; | |||
1064 | model_batch_norm->super.is_trainable = is_trainable; | |||
1065 | ccv_cnnp_model_copy_name(&model_batch_norm->super, name); | |||
1066 | model_batch_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1067 | model_batch_norm->scale.graph = 0; | |||
1068 | model_batch_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1069 | model_batch_norm->bias.graph = 0; | |||
1070 | model_batch_norm->params.bnorm.momentum = momentum; | |||
1071 | model_batch_norm->params.bnorm.epsilon = epsilon; | |||
1072 | return (ccv_cnnp_model_t*)model_batch_norm; | |||
1073 | } | |||
1074 | ||||
1075 | static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1076 | { | |||
1077 | const ccv_cnnp_model_batch_norm_t* const self = (const ccv_cnnp_model_batch_norm_t*)super; | |||
1078 | return ccv_cnnp_batch_norm(self->params.bnorm.momentum, self->params.bnorm.epsilon, self->super.is_trainable, self->super.name); | |||
1079 | } | |||
1080 | ||||
1081 | // MARK - Convolution Layer | |||
1082 | ||||
1083 | typedef struct { | |||
1084 | ccv_cnnp_model_t super; | |||
1085 | ccv_nnc_tensor_symbol_t output; | |||
1086 | ccv_nnc_tensor_symbol_t weights; | |||
1087 | ccv_nnc_tensor_symbol_t bias; | |||
1088 | int groups; | |||
1089 | int filters; | |||
1090 | int kdim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
1091 | int dilation[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
1092 | int no_bias; | |||
1093 | int format; | |||
1094 | ccv_nnc_hint_t hint; | |||
1095 | } ccv_cnnp_model_convolution_t; | |||
1096 | ||||
1097 | static void _ccv_cnnp_convolution_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1098 | { | |||
1099 | ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super; | |||
1100 | PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_convolution_build] -\n"); fflush(stdout); } } while (0); | |||
1101 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1101, __extension__ __PRETTY_FUNCTION__); })); | |||
1102 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1102, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1103 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1104 | int i; | |||
1105 | const int nd = CCV_NNC_MAX_DIM(2) + 2; | |||
1106 | ccv_nnc_tensor_param_t weights_params = params; | |||
1107 | if (self->format) | |||
1108 | weights_params.format = self->format; | |||
1109 | ccv_nnc_tensor_set_n(&weights_params, self->filters); | |||
1110 | assert(ccv_nnc_tensor_get_c(params) % self->groups == 0)((void) sizeof ((ccv_nnc_tensor_get_c(params) % self->groups == 0) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_get_c(params ) % self->groups == 0) ; else __assert_fail ("ccv_nnc_tensor_get_c(params) % self->groups == 0" , "ccv_cnnp_model_addons.c", 1110, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1111 | ccv_nnc_tensor_set_c(&weights_params, nd, ccv_nnc_tensor_get_c(params) / self->groups); | |||
1112 | const int hw = ccv_nnc_tensor_hw(weights_params, nd); | |||
1113 | assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw >= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c" , 1113, __extension__ __PRETTY_FUNCTION__); })); | |||
1114 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1115 | weights_params.dim[i + hw] = self->kdim[i]; | |||
1116 | if (!self->weights.graph) | |||
1117 | self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights"); | |||
1118 | assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__ ({ if (self->weights.graph == graph) ; else __assert_fail ("self->weights.graph == graph", "ccv_cnnp_model_addons.c" , 1118, __extension__ __PRETTY_FUNCTION__); })); | |||
1119 | ccv_nnc_tensor_param_t bias_params = params; | |||
1120 | if (self->format) | |||
1121 | bias_params.format = self->format; | |||
1122 | memset(bias_params.dim, 0, sizeof(bias_params.dim)); | |||
1123 | bias_params.dim[0] = self->filters; | |||
1124 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(self->groups, self->filters)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={}},.convolution={.count=self->filters,.groups =self->groups}}), 0); | |||
1125 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1126 | cmd.info.size.dim[i] = self->kdim[i]; | |||
1127 | memcpy(cmd.info.convolution.dilation, self->dilation, sizeof(self->dilation)); | |||
1128 | ccv_nnc_tensor_param_t output_params; | |||
1129 | // Dilate weight size based on the dilation factor. | |||
1130 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1131 | weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]); typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1; | |||
1132 | ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){ | |||
1133 | params, | |||
1134 | weights_params, | |||
1135 | bias_params, | |||
1136 | }, 3, self->hint, &output_params, 1); | |||
1137 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1138 | ccv_nnc_graph_exec_symbol_t convolution; | |||
1139 | if (self->no_bias) | |||
1140 | convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights }, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution"); | |||
1141 | else { | |||
1142 | if (!self->bias.graph) | |||
1143 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
1144 | convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights, self->bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights , self->bias}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution"); | |||
1145 | } | |||
1146 | ccv_nnc_graph_exec_symbol_set_hint(graph, convolution, self->hint); | |||
1147 | outputs[0] = output; | |||
1148 | } | |||
1149 | ||||
1150 | static void _ccv_cnnp_convolution_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
1151 | { | |||
1152 | ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super; | |||
1153 | const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights); | |||
1154 | const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n (weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b ; }); | |||
1155 | const int count = ccv_nnc_tensor_count(weight_params); | |||
1156 | const float std = sqrtf(2) / sqrtf(count / n); | |||
1157 | const float bound = sqrtf(3) * std; | |||
1158 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
1159 | if (self->bias.graph) | |||
1160 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
1161 | } | |||
1162 | ||||
1163 | static void _ccv_cnnp_convolution_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
1164 | { | |||
1165 | ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super; | |||
1166 | add_to_array(parameters, self->weights, is_trainable); | |||
1167 | if (self->bias.graph) | |||
1168 | add_to_array(parameters, self->bias, is_trainable); | |||
1169 | } | |||
1170 | ||||
1171 | static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1172 | ||||
1173 | static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_isa = { | |||
1174 | .build = _ccv_cnnp_convolution_build, | |||
1175 | .init_states = _ccv_cnnp_convolution_init_states, | |||
1176 | .add_to_parameter = _ccv_cnnp_convolution_add_to_parameter, | |||
1177 | .copy = _ccv_cnnp_convolution_copy, | |||
1178 | }; | |||
1179 | ||||
1180 | ccv_cnnp_model_t* ccv_cnnp_convolution(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name) | |||
1181 | { | |||
1182 | ccv_cnnp_model_convolution_t* const model_convolution = (ccv_cnnp_model_convolution_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_t)); | |||
1183 | model_convolution->super.isa = &ccv_cnnp_convolution_isa; | |||
1184 | model_convolution->super.input_size = 1; | |||
1185 | model_convolution->super.outputs = &model_convolution->output; | |||
1186 | model_convolution->super.output_size = 1; | |||
1187 | model_convolution->super.is_trainable = is_trainable; | |||
1188 | ccv_cnnp_model_copy_name(&model_convolution->super, name); | |||
1189 | model_convolution->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1190 | model_convolution->weights.graph = 0; | |||
1191 | model_convolution->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1192 | model_convolution->bias.graph = 0; | |||
1193 | model_convolution->groups = groups; | |||
1194 | model_convolution->filters = filters; | |||
1195 | memcpy(model_convolution->kdim, kdim, sizeof(model_convolution->kdim)); | |||
1196 | memcpy(model_convolution->dilation, dilation, sizeof(model_convolution->dilation)); | |||
1197 | model_convolution->no_bias = no_bias; | |||
1198 | model_convolution->hint = hint; | |||
1199 | model_convolution->format = format; | |||
1200 | return (ccv_cnnp_model_t*)model_convolution; | |||
1201 | } | |||
1202 | ||||
1203 | static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1204 | { | |||
1205 | ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super; | |||
1206 | return ccv_cnnp_convolution(self->groups, self->filters, self->kdim, self->dilation, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name); | |||
1207 | } | |||
1208 | ||||
1209 | // MARK - Convolution Transpose Layer | |||
1210 | ||||
1211 | typedef struct { | |||
1212 | ccv_cnnp_model_t super; | |||
1213 | ccv_nnc_tensor_symbol_t output; | |||
1214 | ccv_nnc_tensor_symbol_t weights; | |||
1215 | ccv_nnc_tensor_symbol_t bias; | |||
1216 | int groups; | |||
1217 | int filters; | |||
1218 | int kdim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
1219 | int dilation[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
1220 | int output_padding; | |||
1221 | int no_bias; | |||
1222 | int format; | |||
1223 | ccv_nnc_hint_t hint; | |||
1224 | } ccv_cnnp_model_convolution_transpose_t; | |||
1225 | ||||
1226 | static void _ccv_cnnp_convolution_transpose_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1227 | { | |||
1228 | ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super; | |||
1229 | PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_transpose_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_convolution_transpose_build] -\n"); fflush(stdout ); } } while (0); | |||
1230 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1230, __extension__ __PRETTY_FUNCTION__); })); | |||
1231 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1231, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1232 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1233 | int i; | |||
1234 | const int nd = CCV_NNC_MAX_DIM(2) + 2; | |||
1235 | ccv_nnc_tensor_param_t weights_params = params; | |||
1236 | if (self->format) | |||
1237 | weights_params.format = self->format; | |||
1238 | ccv_nnc_tensor_set_n(&weights_params, ccv_nnc_tensor_get_c(params)); | |||
1239 | assert(ccv_nnc_tensor_get_c(params) % self->groups == 0)((void) sizeof ((ccv_nnc_tensor_get_c(params) % self->groups == 0) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_get_c(params ) % self->groups == 0) ; else __assert_fail ("ccv_nnc_tensor_get_c(params) % self->groups == 0" , "ccv_cnnp_model_addons.c", 1239, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1240 | ccv_nnc_tensor_set_c(&weights_params, nd, self->filters / self->groups); | |||
1241 | const int hw = ccv_nnc_tensor_hw(weights_params, nd); | |||
1242 | assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw >= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c" , 1242, __extension__ __PRETTY_FUNCTION__); })); | |||
1243 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1244 | weights_params.dim[i + hw] = self->kdim[i]; | |||
1245 | if (!self->weights.graph) | |||
1246 | self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights"); | |||
1247 | assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__ ({ if (self->weights.graph == graph) ; else __assert_fail ("self->weights.graph == graph", "ccv_cnnp_model_addons.c" , 1247, __extension__ __PRETTY_FUNCTION__); })); | |||
1248 | ccv_nnc_tensor_param_t bias_params = params; | |||
1249 | if (self->format) | |||
1250 | bias_params.format = self->format; | |||
1251 | memset(bias_params.dim, 0, sizeof(bias_params.dim)); | |||
1252 | bias_params.dim[0] = self->filters; | |||
1253 | ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(self->groups, self->filters, self->output_padding)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={}},.convolution_transpose={.count=self->filters ,.groups=self->groups,.output_padding=self->output_padding }}), 0); | |||
1254 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1255 | cmd.info.size.dim[i] = self->kdim[i]; | |||
1256 | memcpy(cmd.info.convolution_transpose.dilation, self->dilation, sizeof(self->dilation)); | |||
1257 | ccv_nnc_tensor_param_t output_params; | |||
1258 | // Dilate weight size based on the dilation factor. | |||
1259 | for (i = 0; i < CCV_NNC_MAX_DIM(2); i++) | |||
1260 | weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]); typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1; | |||
1261 | ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){ | |||
1262 | params, | |||
1263 | weights_params, | |||
1264 | bias_params, | |||
1265 | }, 3, self->hint, &output_params, 1); | |||
1266 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1267 | ccv_nnc_graph_exec_symbol_t convolution_transpose; | |||
1268 | if (self->no_bias) | |||
1269 | convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights }, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose"); | |||
1270 | else { | |||
1271 | if (!self->bias.graph) | |||
1272 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
1273 | convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights, self->bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights , self->bias}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose"); | |||
1274 | } | |||
1275 | ccv_nnc_graph_exec_symbol_set_hint(graph, convolution_transpose, self->hint); | |||
1276 | outputs[0] = output; | |||
1277 | } | |||
1278 | ||||
1279 | static void _ccv_cnnp_convolution_transpose_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
1280 | { | |||
1281 | ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super; | |||
1282 | const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights); | |||
1283 | const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n (weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b ; }); | |||
1284 | const int count = ccv_nnc_tensor_count(weight_params); | |||
1285 | const float std = sqrtf(2) / sqrtf(count / n); | |||
1286 | const float bound = sqrtf(3) * std; | |||
1287 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
1288 | if (self->bias.graph) | |||
1289 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
1290 | } | |||
1291 | ||||
1292 | static void _ccv_cnnp_convolution_transpose_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
1293 | { | |||
1294 | ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super; | |||
1295 | add_to_array(parameters, self->weights, is_trainable); | |||
1296 | if (self->bias.graph) | |||
1297 | add_to_array(parameters, self->bias, is_trainable); | |||
1298 | } | |||
1299 | ||||
1300 | static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1301 | ||||
1302 | static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_transpose_isa = { | |||
1303 | .build = _ccv_cnnp_convolution_transpose_build, | |||
1304 | .init_states = _ccv_cnnp_convolution_transpose_init_states, | |||
1305 | .add_to_parameter = _ccv_cnnp_convolution_transpose_add_to_parameter, | |||
1306 | .copy = _ccv_cnnp_convolution_transpose_copy, | |||
1307 | }; | |||
1308 | ||||
1309 | ccv_cnnp_model_t* ccv_cnnp_convolution_transpose(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int output_padding, const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name) | |||
1310 | { | |||
1311 | ccv_cnnp_model_convolution_transpose_t* const model_convolution_transpose = (ccv_cnnp_model_convolution_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_transpose_t)); | |||
1312 | model_convolution_transpose->super.isa = &ccv_cnnp_convolution_transpose_isa; | |||
1313 | model_convolution_transpose->super.input_size = 1; | |||
1314 | model_convolution_transpose->super.outputs = &model_convolution_transpose->output; | |||
1315 | model_convolution_transpose->super.output_size = 1; | |||
1316 | model_convolution_transpose->super.is_trainable = is_trainable; | |||
1317 | ccv_cnnp_model_copy_name(&model_convolution_transpose->super, name); | |||
1318 | model_convolution_transpose->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1319 | model_convolution_transpose->weights.graph = 0; | |||
1320 | model_convolution_transpose->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1321 | model_convolution_transpose->bias.graph = 0; | |||
1322 | model_convolution_transpose->groups = groups; | |||
1323 | model_convolution_transpose->filters = filters; | |||
1324 | memcpy(model_convolution_transpose->kdim, kdim, sizeof(model_convolution_transpose->kdim)); | |||
1325 | memcpy(model_convolution_transpose->dilation, dilation, sizeof(model_convolution_transpose->dilation)); | |||
1326 | model_convolution_transpose->output_padding = output_padding; | |||
1327 | model_convolution_transpose->no_bias = no_bias; | |||
1328 | model_convolution_transpose->hint = hint; | |||
1329 | model_convolution_transpose->format = format; | |||
1330 | return (ccv_cnnp_model_t*)model_convolution_transpose; | |||
1331 | } | |||
1332 | ||||
1333 | static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1334 | { | |||
1335 | ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super; | |||
1336 | return ccv_cnnp_convolution_transpose(self->groups, self->filters, self->kdim, self->dilation, self->output_padding, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name); | |||
1337 | } | |||
1338 | ||||
1339 | // MARK - Dense Layer | |||
1340 | ||||
1341 | typedef struct { | |||
1342 | ccv_cnnp_model_t super; | |||
1343 | ccv_nnc_tensor_symbol_t output; | |||
1344 | ccv_nnc_tensor_symbol_t weights; | |||
1345 | ccv_nnc_tensor_symbol_t bias; | |||
1346 | int count; | |||
1347 | int no_bias; | |||
1348 | } ccv_cnnp_model_dense_t; | |||
1349 | ||||
1350 | static void _ccv_cnnp_dense_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1351 | { | |||
1352 | ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super; | |||
1353 | PRINT(CCV_CLI_VERBOSE, "[cnnp_dense_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_dense_build] -\n"); fflush(stdout); } } while (0); | |||
1354 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1354, __extension__ __PRETTY_FUNCTION__); })); | |||
1355 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1355, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1356 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1357 | ccv_nnc_tensor_param_t weights_params = params; | |||
1358 | memset(weights_params.dim, 0, sizeof(weights_params.dim)); | |||
1359 | weights_params.dim[0] = self->count; | |||
1360 | weights_params.dim[1] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1]; | |||
1361 | if (!self->weights.graph) | |||
1362 | self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights"); | |||
1363 | assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__ ({ if (self->weights.graph == graph) ; else __assert_fail ("self->weights.graph == graph", "ccv_cnnp_model_addons.c" , 1363, __extension__ __PRETTY_FUNCTION__); })); | |||
1364 | ccv_nnc_tensor_param_t bias_params = params; | |||
1365 | memset(bias_params.dim, 0, sizeof(bias_params.dim)); | |||
1366 | bias_params.dim[0] = self->count; | |||
1367 | ccv_nnc_cmd_t cmd = {0}; | |||
1368 | cmd.cmd = CCV_NNC_GEMM_FORWARD; | |||
1369 | cmd.info.blas.a[0] = 1; | |||
1370 | cmd.info.blas.a[1] = 1; | |||
1371 | cmd.info.blas.transpose_b[0] = 0; | |||
1372 | cmd.info.blas.transpose_b[1] = 1; | |||
1373 | ccv_nnc_tensor_param_t output_params; | |||
1374 | ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){ | |||
1375 | params, | |||
1376 | weights_params, | |||
1377 | bias_params, | |||
1378 | }, 3, ccv_nnc_no_hint, &output_params, 1); | |||
1379 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1380 | if (self->no_bias) | |||
1381 | ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights }, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense"); | |||
1382 | else { | |||
1383 | if (!self->bias.graph) | |||
1384 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
1385 | ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], self->weights, self->bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->weights , self->bias}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense"); | |||
1386 | } | |||
1387 | outputs[0] = output; | |||
1388 | } | |||
1389 | ||||
1390 | static void _ccv_cnnp_dense_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
1391 | { | |||
1392 | ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super; | |||
1393 | const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights); | |||
1394 | const int c = weight_params.dim[1]; | |||
1395 | const float std = sqrtf(2) / sqrtf(c); | |||
1396 | const float bound = sqrtf(3) * std; | |||
1397 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
1398 | if (self->bias.graph) | |||
1399 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
1400 | } | |||
1401 | ||||
1402 | static void _ccv_cnnp_dense_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
1403 | { | |||
1404 | ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super; | |||
1405 | add_to_array(parameters, self->weights, is_trainable); | |||
1406 | if (self->bias.graph) | |||
1407 | add_to_array(parameters, self->bias, is_trainable); | |||
1408 | } | |||
1409 | ||||
1410 | static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1411 | ||||
1412 | static const ccv_cnnp_model_vtab_t ccv_cnnp_dense_isa = { | |||
1413 | .build = _ccv_cnnp_dense_build, | |||
1414 | .init_states = _ccv_cnnp_dense_init_states, | |||
1415 | .add_to_parameter = _ccv_cnnp_dense_add_to_parameter, | |||
1416 | .copy = _ccv_cnnp_dense_copy, | |||
1417 | }; | |||
1418 | ||||
1419 | ccv_cnnp_model_t* ccv_cnnp_dense(const int count, const int no_bias, const int is_trainable, const char* const name) | |||
1420 | { | |||
1421 | ccv_cnnp_model_dense_t* const model_dense = (ccv_cnnp_model_dense_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dense_t)); | |||
1422 | model_dense->super.isa = &ccv_cnnp_dense_isa; | |||
1423 | model_dense->super.input_size = 1; | |||
1424 | model_dense->super.outputs = &model_dense->output; | |||
1425 | model_dense->super.output_size = 1; | |||
1426 | model_dense->super.is_trainable = is_trainable; | |||
1427 | ccv_cnnp_model_copy_name(&model_dense->super, name); | |||
1428 | model_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1429 | model_dense->weights.graph = 0; | |||
1430 | model_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
1431 | model_dense->bias.graph = 0; | |||
1432 | model_dense->count = count; | |||
1433 | model_dense->no_bias = no_bias; | |||
1434 | return (ccv_cnnp_model_t*)model_dense; | |||
1435 | } | |||
1436 | ||||
1437 | static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1438 | { | |||
1439 | const ccv_cnnp_model_dense_t* const self = (const ccv_cnnp_model_dense_t*)super; | |||
1440 | return ccv_cnnp_dense(self->count, self->no_bias, self->super.is_trainable, self->super.name); | |||
1441 | } | |||
1442 | ||||
1443 | // MARK - Pool Layers | |||
1444 | ||||
1445 | typedef struct { | |||
1446 | ccv_cnnp_model_t super; | |||
1447 | ccv_nnc_tensor_symbol_t output; | |||
1448 | int kdim[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
1449 | ccv_nnc_hint_t hint; | |||
1450 | } ccv_cnnp_model_pool_t; | |||
1451 | ||||
1452 | static void _ccv_cnnp_max_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1453 | { | |||
1454 | ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super; | |||
1455 | PRINT(CCV_CLI_VERBOSE, "[cnnp_max_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_max_pool_build] -\n"); fflush(stdout); } } while (0); | |||
1456 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1456, __extension__ __PRETTY_FUNCTION__); })); | |||
1457 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1457, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1458 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1459 | const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim)); | |||
1460 | ccv_nnc_cmd_t cmd; | |||
1461 | if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0) | |||
1462 | cmd = CMD_MAX_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0); | |||
1463 | else | |||
1464 | cmd = CMD_MAX_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0); | |||
1465 | ccv_nnc_tensor_param_t output_params; | |||
1466 | ccv_nnc_hint_tensor_auto(cmd, ¶ms, 1, self->hint, &output_params, 1); | |||
1467 | const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1468 | const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "max_pool"); | |||
1469 | ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint); | |||
1470 | outputs[0] = pool_output; | |||
1471 | } | |||
1472 | ||||
1473 | static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1474 | ||||
1475 | static const ccv_cnnp_model_vtab_t ccv_cnnp_max_pool_isa = { | |||
1476 | .build = _ccv_cnnp_max_pool_build, | |||
1477 | .copy = _ccv_cnnp_max_pool_copy, | |||
1478 | }; | |||
1479 | ||||
1480 | ccv_cnnp_model_t* ccv_cnnp_max_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name) | |||
1481 | { | |||
1482 | ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t)); | |||
1483 | model_pool->super.isa = &ccv_cnnp_max_pool_isa; | |||
1484 | model_pool->super.input_size = 1; | |||
1485 | model_pool->super.outputs = &model_pool->output; | |||
1486 | model_pool->super.output_size = 1; | |||
1487 | ccv_cnnp_model_copy_name(&model_pool->super, name); | |||
1488 | memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim)); | |||
1489 | model_pool->hint = hint; | |||
1490 | return (ccv_cnnp_model_t*)model_pool; | |||
1491 | } | |||
1492 | ||||
1493 | static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1494 | { | |||
1495 | const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super; | |||
1496 | return ccv_cnnp_max_pool(self->kdim, self->hint, self->super.name); | |||
1497 | } | |||
1498 | ||||
1499 | static void _ccv_cnnp_average_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1500 | { | |||
1501 | ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super; | |||
1502 | PRINT(CCV_CLI_VERBOSE, "[cnnp_average_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_average_pool_build] -\n"); fflush(stdout); } } while (0); | |||
1503 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1503, __extension__ __PRETTY_FUNCTION__); })); | |||
1504 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1504, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1505 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1506 | const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim)); | |||
1507 | ccv_nnc_cmd_t cmd; | |||
1508 | if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0) | |||
1509 | cmd = CMD_AVERAGE_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0); | |||
1510 | else | |||
1511 | cmd = CMD_AVERAGE_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0); | |||
1512 | ccv_nnc_tensor_param_t output_params; | |||
1513 | ccv_nnc_hint_tensor_auto(cmd, ¶ms, 1, self->hint, &output_params, 1); | |||
1514 | const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1515 | const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "average_pool"); | |||
1516 | ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint); | |||
1517 | outputs[0] = pool_output; | |||
1518 | } | |||
1519 | ||||
1520 | static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
1521 | ||||
1522 | static const ccv_cnnp_model_vtab_t ccv_cnnp_average_pool_isa = { | |||
1523 | .build = _ccv_cnnp_average_pool_build, | |||
1524 | .copy = _ccv_cnnp_average_pool_copy, | |||
1525 | }; | |||
1526 | ||||
1527 | ccv_cnnp_model_t* ccv_cnnp_average_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name) | |||
1528 | { | |||
1529 | ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t)); | |||
1530 | model_pool->super.isa = &ccv_cnnp_average_pool_isa; | |||
1531 | model_pool->super.input_size = 1; | |||
1532 | model_pool->super.outputs = &model_pool->output; | |||
1533 | model_pool->super.output_size = 1; | |||
1534 | ccv_cnnp_model_copy_name(&model_pool->super, name); | |||
1535 | memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim)); | |||
1536 | model_pool->hint = hint; | |||
1537 | return (ccv_cnnp_model_t*)model_pool; | |||
1538 | } | |||
1539 | ||||
1540 | static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1541 | { | |||
1542 | const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super; | |||
1543 | return ccv_cnnp_average_pool(self->kdim, self->hint, self->super.name); | |||
1544 | } | |||
1545 | ||||
1546 | // MARK - RELU Layer | |||
1547 | ||||
1548 | typedef struct { | |||
1549 | ccv_cnnp_model_t super; | |||
1550 | ccv_nnc_tensor_symbol_t output; | |||
1551 | } ccv_cnnp_model_relu_t; | |||
1552 | ||||
1553 | static void _ccv_cnnp_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1554 | { | |||
1555 | PRINT(CCV_CLI_VERBOSE, "[cnnp_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_relu_build] -\n"); fflush(stdout); } } while (0); | |||
1556 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1556, __extension__ __PRETTY_FUNCTION__); })); | |||
1557 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1557, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1558 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1559 | ccv_nnc_tensor_param_t output_params; | |||
1560 | const ccv_nnc_cmd_t relu = CMD_RELU_FORWARD()ccv_nnc_cmd(CCV_NNC_RELU_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
1561 | ccv_nnc_hint_tensor_auto(relu, (ccv_nnc_tensor_param_t []){ | |||
1562 | params, | |||
1563 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1564 | const ccv_nnc_tensor_symbol_t relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1565 | ccv_nnc_graph_exec_symbol_new(graph, relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(relu_output)(const ccv_nnc_tensor_symbol_t []){relu_output}, (1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "relu"); | |||
1566 | outputs[0] = relu_output; | |||
1567 | } | |||
1568 | ||||
1569 | static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1570 | ||||
1571 | static const ccv_cnnp_model_vtab_t ccv_cnnp_relu_isa = { | |||
1572 | .build = _ccv_cnnp_relu_build, | |||
1573 | .copy = _ccv_cnnp_relu_copy, | |||
1574 | }; | |||
1575 | ||||
1576 | ccv_cnnp_model_t* ccv_cnnp_relu(const char* const name) | |||
1577 | { | |||
1578 | ccv_cnnp_model_relu_t* const model_relu = (ccv_cnnp_model_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_relu_t)); | |||
1579 | model_relu->super.isa = &ccv_cnnp_relu_isa; | |||
1580 | model_relu->super.input_size = 1; | |||
1581 | model_relu->super.outputs = &model_relu->output; | |||
1582 | model_relu->super.output_size = 1; | |||
1583 | ccv_cnnp_model_copy_name(&model_relu->super, name); | |||
1584 | return (ccv_cnnp_model_t*)model_relu; | |||
1585 | } | |||
1586 | ||||
1587 | static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
1588 | { | |||
1589 | return ccv_cnnp_relu(self->name); | |||
1590 | } | |||
1591 | ||||
1592 | // MARK - Sigmoid Layer | |||
1593 | ||||
1594 | typedef struct { | |||
1595 | ccv_cnnp_model_t super; | |||
1596 | ccv_nnc_tensor_symbol_t output; | |||
1597 | } ccv_cnnp_model_sigmoid_t; | |||
1598 | ||||
1599 | static void _ccv_cnnp_sigmoid_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1600 | { | |||
1601 | PRINT(CCV_CLI_VERBOSE, "[cnnp_sigmoid_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_sigmoid_build] -\n"); fflush(stdout); } } while (0); | |||
1602 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1602, __extension__ __PRETTY_FUNCTION__); })); | |||
1603 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1603, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1604 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1605 | ccv_nnc_tensor_param_t output_params; | |||
1606 | const ccv_nnc_cmd_t sigmoid = CMD_SIGMOID_FORWARD()ccv_nnc_cmd(CCV_NNC_SIGMOID_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
1607 | ccv_nnc_hint_tensor_auto(sigmoid, (ccv_nnc_tensor_param_t []){ | |||
1608 | params, | |||
1609 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1610 | const ccv_nnc_tensor_symbol_t sigmoid_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1611 | ccv_nnc_graph_exec_symbol_new(graph, sigmoid, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(sigmoid_output)(const ccv_nnc_tensor_symbol_t []){sigmoid_output}, (1 +1 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1 ), "sigmoid"); | |||
1612 | outputs[0] = sigmoid_output; | |||
1613 | } | |||
1614 | ||||
1615 | static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1616 | ||||
1617 | static const ccv_cnnp_model_vtab_t ccv_cnnp_sigmoid_isa = { | |||
1618 | .build = _ccv_cnnp_sigmoid_build, | |||
1619 | .copy = _ccv_cnnp_sigmoid_copy, | |||
1620 | }; | |||
1621 | ||||
1622 | ccv_cnnp_model_t* ccv_cnnp_sigmoid(const char* const name) | |||
1623 | { | |||
1624 | ccv_cnnp_model_sigmoid_t* const model_sigmoid = (ccv_cnnp_model_sigmoid_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sigmoid_t)); | |||
1625 | model_sigmoid->super.isa = &ccv_cnnp_sigmoid_isa; | |||
1626 | model_sigmoid->super.input_size = 1; | |||
1627 | model_sigmoid->super.outputs = &model_sigmoid->output; | |||
1628 | model_sigmoid->super.output_size = 1; | |||
1629 | ccv_cnnp_model_copy_name(&model_sigmoid->super, name); | |||
1630 | return (ccv_cnnp_model_t*)model_sigmoid; | |||
1631 | } | |||
1632 | ||||
1633 | static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
1634 | { | |||
1635 | return ccv_cnnp_sigmoid(self->name); | |||
1636 | } | |||
1637 | ||||
1638 | // MARK - Tanh Layer | |||
1639 | ||||
1640 | typedef struct { | |||
1641 | ccv_cnnp_model_t super; | |||
1642 | ccv_nnc_tensor_symbol_t output; | |||
1643 | } ccv_cnnp_model_tanh_t; | |||
1644 | ||||
1645 | static void _ccv_cnnp_tanh_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1646 | { | |||
1647 | PRINT(CCV_CLI_VERBOSE, "[cnnp_tanh_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_tanh_build] -\n"); fflush(stdout); } } while (0); | |||
1648 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1648, __extension__ __PRETTY_FUNCTION__); })); | |||
1649 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1649, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1650 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1651 | ccv_nnc_tensor_param_t output_params; | |||
1652 | const ccv_nnc_cmd_t tanh = CMD_TANH_FORWARD()ccv_nnc_cmd(CCV_NNC_TANH_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
1653 | ccv_nnc_hint_tensor_auto(tanh, (ccv_nnc_tensor_param_t []){ | |||
1654 | params, | |||
1655 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1656 | const ccv_nnc_tensor_symbol_t tanh_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1657 | ccv_nnc_graph_exec_symbol_new(graph, tanh, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(tanh_output)(const ccv_nnc_tensor_symbol_t []){tanh_output}, (1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "tanh"); | |||
1658 | outputs[0] = tanh_output; | |||
1659 | } | |||
1660 | ||||
1661 | static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1662 | ||||
1663 | static const ccv_cnnp_model_vtab_t ccv_cnnp_tanh_isa = { | |||
1664 | .build = _ccv_cnnp_tanh_build, | |||
1665 | .copy = _ccv_cnnp_tanh_copy, | |||
1666 | }; | |||
1667 | ||||
1668 | ccv_cnnp_model_t* ccv_cnnp_tanh(const char* const name) | |||
1669 | { | |||
1670 | ccv_cnnp_model_tanh_t* const model_tanh = (ccv_cnnp_model_tanh_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_tanh_t)); | |||
1671 | model_tanh->super.isa = &ccv_cnnp_tanh_isa; | |||
1672 | model_tanh->super.input_size = 1; | |||
1673 | model_tanh->super.outputs = &model_tanh->output; | |||
1674 | model_tanh->super.output_size = 1; | |||
1675 | ccv_cnnp_model_copy_name(&model_tanh->super, name); | |||
1676 | return (ccv_cnnp_model_t*)model_tanh; | |||
1677 | } | |||
1678 | ||||
1679 | static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
1680 | { | |||
1681 | return ccv_cnnp_tanh(self->name); | |||
1682 | } | |||
1683 | ||||
1684 | // MARK - Swish Layer | |||
1685 | ||||
1686 | typedef struct { | |||
1687 | ccv_cnnp_model_t super; | |||
1688 | ccv_nnc_tensor_symbol_t output; | |||
1689 | } ccv_cnnp_model_swish_t; | |||
1690 | ||||
1691 | static void _ccv_cnnp_swish_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1692 | { | |||
1693 | PRINT(CCV_CLI_VERBOSE, "[cnnp_swish_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_swish_build] -\n"); fflush(stdout); } } while (0); | |||
1694 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1694, __extension__ __PRETTY_FUNCTION__); })); | |||
1695 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1695, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1696 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1697 | ccv_nnc_tensor_param_t output_params; | |||
1698 | const ccv_nnc_cmd_t swish = CMD_SWISH_FORWARD()ccv_nnc_cmd(CCV_NNC_SWISH_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
1699 | ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){ | |||
1700 | params, | |||
1701 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1702 | const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1703 | ccv_nnc_graph_exec_symbol_new(graph, swish, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(swish_output)(const ccv_nnc_tensor_symbol_t []){swish_output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "swish"); | |||
1704 | outputs[0] = swish_output; | |||
1705 | } | |||
1706 | ||||
1707 | static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1708 | ||||
1709 | static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = { | |||
1710 | .build = _ccv_cnnp_swish_build, | |||
1711 | .copy = _ccv_cnnp_swish_copy, | |||
1712 | }; | |||
1713 | ||||
1714 | ccv_cnnp_model_t* ccv_cnnp_swish(const char* const name) | |||
1715 | { | |||
1716 | ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t)); | |||
1717 | model_swish->super.isa = &ccv_cnnp_swish_isa; | |||
1718 | model_swish->super.input_size = 1; | |||
1719 | model_swish->super.outputs = &model_swish->output; | |||
1720 | model_swish->super.output_size = 1; | |||
1721 | ccv_cnnp_model_copy_name(&model_swish->super, name); | |||
1722 | return (ccv_cnnp_model_t*)model_swish; | |||
1723 | } | |||
1724 | ||||
1725 | static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
1726 | { | |||
1727 | return ccv_cnnp_swish(self->name); | |||
1728 | } | |||
1729 | ||||
1730 | // MARK - GELU Layer | |||
1731 | ||||
1732 | typedef struct { | |||
1733 | ccv_cnnp_model_t super; | |||
1734 | ccv_nnc_tensor_symbol_t output; | |||
1735 | int tanh; | |||
1736 | } ccv_cnnp_model_gelu_t; | |||
1737 | ||||
1738 | static void _ccv_cnnp_gelu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1739 | { | |||
1740 | PRINT(CCV_CLI_VERBOSE, "[cnnp_gelu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_gelu_build] -\n"); fflush(stdout); } } while (0); | |||
1741 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1741, __extension__ __PRETTY_FUNCTION__); })); | |||
1742 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1742, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1743 | ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super; | |||
1744 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1745 | ccv_nnc_tensor_param_t output_params; | |||
1746 | const ccv_nnc_cmd_t gelu = CMD_GELU_FORWARD(self->tanh)ccv_nnc_cmd(CCV_NNC_GELU_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.gelu={.tanh=self->tanh}}, 0); | |||
1747 | ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){ | |||
1748 | params, | |||
1749 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1750 | const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1751 | ccv_nnc_graph_exec_symbol_new(graph, gelu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(gelu_output)(const ccv_nnc_tensor_symbol_t []){gelu_output}, (1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "gelu"); | |||
1752 | outputs[0] = gelu_output; | |||
1753 | } | |||
1754 | ||||
1755 | static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1756 | ||||
1757 | static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = { | |||
1758 | .build = _ccv_cnnp_gelu_build, | |||
1759 | .copy = _ccv_cnnp_gelu_copy, | |||
1760 | }; | |||
1761 | ||||
1762 | ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name) | |||
1763 | { | |||
1764 | ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t)); | |||
1765 | model_gelu->super.isa = &ccv_cnnp_gelu_isa; | |||
1766 | model_gelu->super.input_size = 1; | |||
1767 | model_gelu->super.outputs = &model_gelu->output; | |||
1768 | model_gelu->super.output_size = 1; | |||
1769 | model_gelu->tanh = tanh; | |||
1770 | ccv_cnnp_model_copy_name(&model_gelu->super, name); | |||
1771 | return (ccv_cnnp_model_t*)model_gelu; | |||
1772 | } | |||
1773 | ||||
1774 | static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1775 | { | |||
1776 | ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super; | |||
1777 | return ccv_cnnp_gelu(self->tanh, self->super.name); | |||
1778 | } | |||
1779 | ||||
1780 | // MARK - Leaky ReLU Layer | |||
1781 | ||||
1782 | typedef struct { | |||
1783 | ccv_cnnp_model_t super; | |||
1784 | ccv_nnc_tensor_symbol_t output; | |||
1785 | float negative_slope; | |||
1786 | } ccv_cnnp_model_leaky_relu_t; | |||
1787 | ||||
1788 | static void _ccv_cnnp_leaky_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1789 | { | |||
1790 | PRINT(CCV_CLI_VERBOSE, "[cnnp_leaky_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_leaky_relu_build] -\n"); fflush(stdout); } } while (0); | |||
1791 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1791, __extension__ __PRETTY_FUNCTION__); })); | |||
1792 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1792, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1793 | ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super; | |||
1794 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1795 | ccv_nnc_tensor_param_t output_params; | |||
1796 | const ccv_nnc_cmd_t leaky_relu = CMD_LEAKY_RELU_FORWARD(self->negative_slope)ccv_nnc_cmd(CCV_NNC_LEAKY_RELU_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.leaky_relu={.negative_slope=self-> negative_slope}}, 0); | |||
1797 | ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){ | |||
1798 | params, | |||
1799 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1800 | const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1801 | ccv_nnc_graph_exec_symbol_new(graph, leaky_relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(leaky_relu_output)(const ccv_nnc_tensor_symbol_t []){leaky_relu_output}, (1 +1 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "leaky_relu"); | |||
1802 | outputs[0] = leaky_relu_output; | |||
1803 | } | |||
1804 | ||||
1805 | static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1806 | ||||
1807 | static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = { | |||
1808 | .build = _ccv_cnnp_leaky_relu_build, | |||
1809 | .copy = _ccv_cnnp_leaky_relu_copy, | |||
1810 | }; | |||
1811 | ||||
1812 | ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name) | |||
1813 | { | |||
1814 | ccv_cnnp_model_leaky_relu_t* const model_leaky_relu = (ccv_cnnp_model_leaky_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_leaky_relu_t)); | |||
1815 | model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa; | |||
1816 | model_leaky_relu->super.input_size = 1; | |||
1817 | model_leaky_relu->super.outputs = &model_leaky_relu->output; | |||
1818 | model_leaky_relu->super.output_size = 1; | |||
1819 | model_leaky_relu->negative_slope = negative_slope; | |||
1820 | ccv_cnnp_model_copy_name(&model_leaky_relu->super, name); | |||
1821 | return (ccv_cnnp_model_t*)model_leaky_relu; | |||
1822 | } | |||
1823 | ||||
1824 | static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1825 | { | |||
1826 | ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super; | |||
1827 | return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name); | |||
1828 | } | |||
1829 | ||||
1830 | // MARK - Softmax Layer | |||
1831 | ||||
1832 | typedef struct { | |||
1833 | ccv_cnnp_model_t super; | |||
1834 | ccv_nnc_tensor_symbol_t output; | |||
1835 | } ccv_cnnp_model_softmax_t; | |||
1836 | ||||
1837 | static void _ccv_cnnp_softmax_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1838 | { | |||
1839 | PRINT(CCV_CLI_VERBOSE, "[cnnp_softmax_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_softmax_build] -\n"); fflush(stdout); } } while (0); | |||
1840 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1840, __extension__ __PRETTY_FUNCTION__); })); | |||
1841 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1841, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1842 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1843 | ccv_nnc_tensor_param_t output_params; | |||
1844 | const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
1845 | ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){ | |||
1846 | params, | |||
1847 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1848 | const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1849 | ccv_nnc_graph_exec_symbol_new(graph, softmax, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(softmax_output)(const ccv_nnc_tensor_symbol_t []){softmax_output}, (1 +1 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1 ), "softmax"); | |||
1850 | outputs[0] = softmax_output; | |||
1851 | } | |||
1852 | ||||
1853 | static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1854 | ||||
1855 | static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = { | |||
1856 | .build = _ccv_cnnp_softmax_build, | |||
1857 | .copy = _ccv_cnnp_softmax_copy, | |||
1858 | }; | |||
1859 | ||||
1860 | ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name) | |||
1861 | { | |||
1862 | ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t)); | |||
1863 | model_softmax->super.isa = &ccv_cnnp_softmax_isa; | |||
1864 | model_softmax->super.input_size = 1; | |||
1865 | model_softmax->super.outputs = &model_softmax->output; | |||
1866 | model_softmax->super.output_size = 1; | |||
1867 | ccv_cnnp_model_copy_name(&model_softmax->super, name); | |||
1868 | return (ccv_cnnp_model_t*)model_softmax; | |||
1869 | } | |||
1870 | ||||
1871 | static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context) | |||
1872 | { | |||
1873 | return ccv_cnnp_softmax(self->name); | |||
1874 | } | |||
1875 | ||||
1876 | // MARK - Add Layer | |||
1877 | ||||
1878 | typedef struct { | |||
1879 | ccv_cnnp_model_t super; | |||
1880 | float p; | |||
1881 | float q; | |||
1882 | ccv_nnc_tensor_symbol_t output; | |||
1883 | } ccv_cnnp_model_add_t; | |||
1884 | ||||
1885 | static void _ccv_cnnp_add_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1886 | { | |||
1887 | PRINT(CCV_CLI_VERBOSE, "[cnnp_add_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_add_build] -\n"); fflush(stdout); } } while ( 0); | |||
1888 | const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super; | |||
1889 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 1889, __extension__ __PRETTY_FUNCTION__); })); | |||
1890 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1890, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1891 | ccv_nnc_tensor_param_t input_params[2]; | |||
1892 | int i; | |||
1893 | for (i = 0; i < 2; i++) | |||
1894 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
1895 | ccv_nnc_tensor_param_t output_params; | |||
1896 | const ccv_nnc_cmd_t add = CMD_ADD_FORWARD(self->p, self->q)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={self->p, self->q}}}, 0); | |||
1897 | ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
1898 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1899 | ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add"); | |||
1900 | } | |||
1901 | ||||
1902 | static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1903 | ||||
1904 | static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = { | |||
1905 | .build = _ccv_cnnp_add_build, | |||
1906 | .copy = _ccv_cnnp_add_copy, | |||
1907 | }; | |||
1908 | ||||
1909 | ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name) | |||
1910 | { | |||
1911 | ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t)); | |||
1912 | model_add->super.isa = &ccv_cnnp_add_isa; | |||
1913 | model_add->super.input_size = 2; | |||
1914 | model_add->super.outputs = &model_add->output; | |||
1915 | model_add->super.output_size = 1; | |||
1916 | model_add->p = p; | |||
1917 | model_add->q = q; | |||
1918 | ccv_cnnp_model_copy_name(&model_add->super, name); | |||
1919 | return (ccv_cnnp_model_t*)model_add; | |||
1920 | } | |||
1921 | ||||
1922 | static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1923 | { | |||
1924 | const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super; | |||
1925 | return ccv_cnnp_add(self->p, self->q, self->super.name); | |||
1926 | } | |||
1927 | ||||
1928 | // MARK - Mul Layer | |||
1929 | ||||
1930 | typedef struct { | |||
1931 | ccv_cnnp_model_t super; | |||
1932 | ccv_nnc_tensor_symbol_t output; | |||
1933 | float p; | |||
1934 | } ccv_cnnp_model_mul_t; | |||
1935 | ||||
1936 | static void _ccv_cnnp_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1937 | { | |||
1938 | PRINT(CCV_CLI_VERBOSE, "[cnnp_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_mul_build] -\n"); fflush(stdout); } } while ( 0); | |||
1939 | const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super; | |||
1940 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 1940, __extension__ __PRETTY_FUNCTION__); })); | |||
1941 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1941, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1942 | ccv_nnc_tensor_param_t input_params[2]; | |||
1943 | int i; | |||
1944 | for (i = 0; i < 2; i++) | |||
1945 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
1946 | ccv_nnc_tensor_param_t output_params; | |||
1947 | const ccv_nnc_cmd_t mul = CMD_MUL_FORWARD(self->p)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={self->p,}}}, 0); | |||
1948 | ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
1949 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1950 | ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul"); | |||
1951 | } | |||
1952 | ||||
1953 | static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
1954 | ||||
1955 | static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = { | |||
1956 | .build = _ccv_cnnp_mul_build, | |||
1957 | .copy = _ccv_cnnp_mul_copy, | |||
1958 | }; | |||
1959 | ||||
1960 | ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name) | |||
1961 | { | |||
1962 | ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t)); | |||
1963 | model_mul->super.isa = &ccv_cnnp_mul_isa; | |||
1964 | model_mul->super.input_size = 2; | |||
1965 | model_mul->super.outputs = &model_mul->output; | |||
1966 | model_mul->super.output_size = 1; | |||
1967 | model_mul->p = p; | |||
1968 | ccv_cnnp_model_copy_name(&model_mul->super, name); | |||
1969 | return (ccv_cnnp_model_t*)model_mul; | |||
1970 | } | |||
1971 | ||||
1972 | static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
1973 | { | |||
1974 | const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super; | |||
1975 | return ccv_cnnp_mul(self->p, self->super.name); | |||
1976 | } | |||
1977 | ||||
1978 | // MARK - Scalar Mul Layer | |||
1979 | ||||
1980 | typedef struct { | |||
1981 | ccv_cnnp_model_t super; | |||
1982 | ccv_nnc_tensor_symbol_t output; | |||
1983 | float a; | |||
1984 | } ccv_cnnp_model_scalar_mul_t; | |||
1985 | ||||
1986 | static void _ccv_cnnp_scalar_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
1987 | { | |||
1988 | PRINT(CCV_CLI_VERBOSE, "[cnnp_scalar_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_scalar_mul_build] -\n"); fflush(stdout); } } while (0); | |||
1989 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 1989, __extension__ __PRETTY_FUNCTION__); })); | |||
1990 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 1990, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1991 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
1992 | ccv_nnc_tensor_param_t output_params; | |||
1993 | ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super; | |||
1994 | const ccv_nnc_cmd_t scalar_mul = CMD_SCALAR_MUL_FORWARD(self->a)ccv_nnc_cmd(CCV_NNC_SCALAR_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={self->a,}}}, 0); | |||
1995 | ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){ | |||
1996 | params, | |||
1997 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
1998 | const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
1999 | ccv_nnc_graph_exec_symbol_new(graph, scalar_mul, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(scalar_mul_output)(const ccv_nnc_tensor_symbol_t []){scalar_mul_output}, (1 +1 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "scalar_mul"); | |||
2000 | outputs[0] = scalar_mul_output; | |||
2001 | } | |||
2002 | ||||
2003 | static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2004 | ||||
2005 | static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = { | |||
2006 | .build = _ccv_cnnp_scalar_mul_build, | |||
2007 | .copy = _ccv_cnnp_scalar_mul_copy, | |||
2008 | }; | |||
2009 | ||||
2010 | ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name) | |||
2011 | { | |||
2012 | ccv_cnnp_model_scalar_mul_t* const model_scalar_mul = (ccv_cnnp_model_scalar_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_mul_t)); | |||
2013 | model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa; | |||
2014 | model_scalar_mul->super.input_size = 1; | |||
2015 | model_scalar_mul->super.outputs = &model_scalar_mul->output; | |||
2016 | model_scalar_mul->super.output_size = 1; | |||
2017 | model_scalar_mul->a = a; | |||
2018 | ccv_cnnp_model_copy_name(&model_scalar_mul->super, name); | |||
2019 | return (ccv_cnnp_model_t*)model_scalar_mul; | |||
2020 | } | |||
2021 | ||||
2022 | static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2023 | { | |||
2024 | const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super; | |||
2025 | return ccv_cnnp_scalar_mul(self->a, self->super.name); | |||
2026 | } | |||
2027 | ||||
2028 | // MARK - Div Layer | |||
2029 | ||||
2030 | typedef struct { | |||
2031 | ccv_cnnp_model_t super; | |||
2032 | ccv_nnc_tensor_symbol_t output; | |||
2033 | int reciprocal; | |||
2034 | } ccv_cnnp_model_div_t; | |||
2035 | ||||
2036 | static void _ccv_cnnp_div_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2037 | { | |||
2038 | const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super; | |||
2039 | PRINT(CCV_CLI_VERBOSE, "[cnnp_div_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_div_build] -\n"); fflush(stdout); } } while ( 0); | |||
2040 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2040, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2041 | ccv_nnc_tensor_param_t input_params[2]; | |||
2042 | int i; | |||
2043 | ccv_nnc_tensor_param_t output_params; | |||
2044 | const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
2045 | if (self->reciprocal) | |||
2046 | { | |||
2047 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2047, __extension__ __PRETTY_FUNCTION__); })); | |||
2048 | input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2049 | input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2050 | ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
2051 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2052 | ccv_nnc_graph_exec_symbol_new(graph, div, TENSOR_SYMBOL_LIST(NO_TENSOR_SYMBOL, inputs[0])(const ccv_nnc_tensor_symbol_t []){(const ccv_nnc_tensor_symbol_t ){.d = CCV_NNC_NO_TENSOR_SYMBOL}, inputs[0]}, (1 +1 +1 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), outputs, output_size, "div"); | |||
2053 | } else { | |||
2054 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 2054, __extension__ __PRETTY_FUNCTION__); })); | |||
2055 | for (i = 0; i < 2; i++) | |||
2056 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
2057 | ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1); | |||
2058 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2059 | ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div"); | |||
2060 | } | |||
2061 | } | |||
2062 | ||||
2063 | static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
2064 | ||||
2065 | static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = { | |||
2066 | .build = _ccv_cnnp_div_build, | |||
2067 | .copy = _ccv_cnnp_div_copy, | |||
2068 | }; | |||
2069 | ||||
2070 | ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name) | |||
2071 | { | |||
2072 | ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t)); | |||
2073 | model_div->super.isa = &ccv_cnnp_div_isa; | |||
2074 | model_div->super.input_size = reciprocal ? 1 : 2; | |||
2075 | model_div->super.outputs = &model_div->output; | |||
2076 | model_div->super.output_size = 1; | |||
2077 | model_div->reciprocal = reciprocal; | |||
2078 | ccv_cnnp_model_copy_name(&model_div->super, name); | |||
2079 | return (ccv_cnnp_model_t*)model_div; | |||
2080 | } | |||
2081 | ||||
2082 | static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2083 | { | |||
2084 | const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super; | |||
2085 | return ccv_cnnp_div(self->reciprocal, self->super.name); | |||
2086 | } | |||
2087 | ||||
2088 | // MARK - Sqrt Layer | |||
2089 | ||||
2090 | typedef struct { | |||
2091 | ccv_cnnp_model_t super; | |||
2092 | ccv_nnc_tensor_symbol_t output; | |||
2093 | } ccv_cnnp_model_sqrt_t; | |||
2094 | ||||
2095 | static void _ccv_cnnp_sqrt_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2096 | { | |||
2097 | PRINT(CCV_CLI_VERBOSE, "[cnnp_sqrt_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_sqrt_build] -\n"); fflush(stdout); } } while (0); | |||
2098 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2098, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2099 | ccv_nnc_tensor_param_t input_params[1]; | |||
2100 | ccv_nnc_tensor_param_t output_params; | |||
2101 | const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0); | |||
2102 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2102, __extension__ __PRETTY_FUNCTION__); })); | |||
2103 | input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2104 | ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
2105 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2106 | ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt"); | |||
2107 | } | |||
2108 | ||||
2109 | static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
2110 | ||||
2111 | static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = { | |||
2112 | .build = _ccv_cnnp_sqrt_build, | |||
2113 | .copy = _ccv_cnnp_sqrt_copy, | |||
2114 | }; | |||
2115 | ||||
2116 | ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name) | |||
2117 | { | |||
2118 | ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t)); | |||
2119 | model_sqrt->super.isa = &ccv_cnnp_sqrt_isa; | |||
2120 | model_sqrt->super.input_size = 1; | |||
2121 | model_sqrt->super.outputs = &model_sqrt->output; | |||
2122 | model_sqrt->super.output_size = 1; | |||
2123 | ccv_cnnp_model_copy_name(&model_sqrt->super, name); | |||
2124 | return (ccv_cnnp_model_t*)model_sqrt; | |||
2125 | } | |||
2126 | ||||
2127 | static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2128 | { | |||
2129 | const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super; | |||
2130 | return ccv_cnnp_sqrt(self->super.name); | |||
2131 | } | |||
2132 | ||||
2133 | // MARK - Cmul Layer | |||
2134 | ||||
2135 | typedef struct { | |||
2136 | ccv_cnnp_model_t super; | |||
2137 | ccv_nnc_tensor_symbol_t output; | |||
2138 | } ccv_cnnp_model_cmul_t; | |||
2139 | ||||
2140 | static void _ccv_cnnp_cmul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2141 | { | |||
2142 | PRINT(CCV_CLI_VERBOSE, "[cnnp_cmul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_cmul_build] -\n"); fflush(stdout); } } while (0); | |||
2143 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 2143, __extension__ __PRETTY_FUNCTION__); })); | |||
2144 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2144, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2145 | ccv_nnc_tensor_param_t input_params[2]; | |||
2146 | int i; | |||
2147 | for (i = 0; i < 2; i++) | |||
2148 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
2149 | ccv_nnc_tensor_param_t output_params; | |||
2150 | const ccv_nnc_cmd_t mul = CMD_CMUL_FORWARD()ccv_nnc_cmd(CCV_NNC_CMUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}}}, 0); | |||
2151 | ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
2152 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2153 | ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul"); | |||
2154 | } | |||
2155 | ||||
2156 | static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
2157 | ||||
2158 | static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = { | |||
2159 | .build = _ccv_cnnp_cmul_build, | |||
2160 | .copy = _ccv_cnnp_cmul_copy, | |||
2161 | }; | |||
2162 | ||||
2163 | ccv_cnnp_model_t* ccv_cnnp_cmul(const char* const name) | |||
2164 | { | |||
2165 | ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t)); | |||
2166 | model_cmul->super.isa = &ccv_cnnp_cmul_isa; | |||
2167 | model_cmul->super.input_size = 2; | |||
2168 | model_cmul->super.outputs = &model_cmul->output; | |||
2169 | model_cmul->super.output_size = 1; | |||
2170 | ccv_cnnp_model_copy_name(&model_cmul->super, name); | |||
2171 | return (ccv_cnnp_model_t*)model_cmul; | |||
2172 | } | |||
2173 | ||||
2174 | static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2175 | { | |||
2176 | return ccv_cnnp_cmul(super->name); | |||
2177 | } | |||
2178 | ||||
2179 | // MARK - Transpose Layer | |||
2180 | ||||
2181 | typedef struct { | |||
2182 | ccv_cnnp_model_t super; | |||
2183 | ccv_nnc_tensor_symbol_t output; | |||
2184 | int transpose[2]; | |||
2185 | } ccv_cnnp_model_transpose_t; | |||
2186 | ||||
2187 | static void _ccv_cnnp_transpose_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2188 | { | |||
2189 | ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super; | |||
2190 | PRINT(CCV_CLI_VERBOSE, "[cnnp_transpose_build] (%d, %d)\n", self->transpose[0], self->transpose[1])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_transpose_build] (%d, %d)\n", self->transpose [0], self->transpose[1]); fflush(stdout); } } while (0); | |||
2191 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2191, __extension__ __PRETTY_FUNCTION__); })); | |||
2192 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2192, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2193 | if (self->transpose[0] == self->transpose[1]) | |||
2194 | { | |||
2195 | outputs[0] = inputs[0]; | |||
2196 | return; | |||
2197 | } | |||
2198 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2199 | ccv_nnc_tensor_param_t output_params; | |||
2200 | const ccv_nnc_cmd_t transpose = CMD_TRANSPOSE_FORWARD(self->transpose[0], self->transpose[1])ccv_nnc_cmd(CCV_NNC_TRANSPOSE_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.transpose={.axis={self->transpose[ 0], self->transpose[1]}}}), 0); | |||
2201 | ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){ | |||
2202 | params, | |||
2203 | }, 1, ccv_nnc_no_hint, &output_params, 1); | |||
2204 | const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2205 | ccv_nnc_graph_exec_symbol_new(graph, transpose, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(transpose_output)(const ccv_nnc_tensor_symbol_t []){transpose_output}, (1 +1 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "transpose"); | |||
2206 | outputs[0] = transpose_output; | |||
2207 | } | |||
2208 | ||||
2209 | static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2210 | ||||
2211 | static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = { | |||
2212 | .build = _ccv_cnnp_transpose_build, | |||
2213 | .copy = _ccv_cnnp_transpose_copy, | |||
2214 | }; | |||
2215 | ||||
2216 | ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name) | |||
2217 | { | |||
2218 | ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t)); | |||
2219 | model_transpose->super.isa = &ccv_cnnp_transpose_isa; | |||
2220 | model_transpose->super.input_size = 1; | |||
2221 | model_transpose->super.outputs = &model_transpose->output; | |||
2222 | model_transpose->super.output_size = 1; | |||
2223 | model_transpose->transpose[0] = axis_a; | |||
2224 | model_transpose->transpose[1] = axis_b; | |||
2225 | ccv_cnnp_model_copy_name(&model_transpose->super, name); | |||
2226 | return (ccv_cnnp_model_t*)model_transpose; | |||
2227 | } | |||
2228 | ||||
2229 | static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2230 | { | |||
2231 | const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super; | |||
2232 | return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name); | |||
2233 | } | |||
2234 | ||||
2235 | // MARK - Layer Norm Layer | |||
2236 | ||||
2237 | typedef struct { | |||
2238 | ccv_cnnp_model_t super; | |||
2239 | ccv_nnc_tensor_symbol_t output; | |||
2240 | ccv_nnc_tensor_symbol_t bias; | |||
2241 | ccv_nnc_tensor_symbol_t scale; | |||
2242 | ccv_nnc_cmd_param_t params; | |||
2243 | } ccv_cnnp_model_layer_norm_t; | |||
2244 | ||||
2245 | static void _ccv_cnnp_layer_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2246 | { | |||
2247 | PRINT(CCV_CLI_VERBOSE, "[cnnp_layer_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_layer_norm_build] -\n"); fflush(stdout); } } while (0); | |||
2248 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2248, __extension__ __PRETTY_FUNCTION__); })); | |||
2249 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2249, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2250 | ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super; | |||
2251 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2252 | ccv_nnc_tensor_param_t bias_params = params; | |||
2253 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
2254 | int i; | |||
2255 | for (i = 0; i < nd; i++) | |||
2256 | bias_params.dim[i] = 1; | |||
2257 | for (i = 0; i < self->params.lnorm.count; i++) | |||
2258 | bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]]; | |||
2259 | if (self->params.lnorm.elementwise_affine) | |||
2260 | { | |||
2261 | // Both scale and bias are shared between if this model is reused. | |||
2262 | if (!self->scale.graph) | |||
2263 | self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale"); | |||
2264 | if (!self->bias.graph) | |||
2265 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
2266 | } | |||
2267 | const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0); | |||
2268 | ccv_nnc_tensor_param_t output_params[3]; | |||
2269 | if (self->params.lnorm.elementwise_affine) | |||
2270 | ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){ | |||
2271 | params, | |||
2272 | bias_params, | |||
2273 | bias_params, | |||
2274 | }, 3, ccv_nnc_no_hint, output_params, 3); | |||
2275 | else | |||
2276 | ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){ | |||
2277 | params, | |||
2278 | }, 1, ccv_nnc_no_hint, output_params, 3); | |||
2279 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
2280 | const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean"); | |||
2281 | const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std"); | |||
2282 | if (self->params.lnorm.elementwise_affine) | |||
2283 | ccv_nnc_graph_exec_symbol_new(graph, layer_norm, TENSOR_SYMBOL_LIST(inputs[0], self->scale, self->bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->scale, self->bias}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std }, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), "layer_norm"); | |||
2284 | else | |||
2285 | ccv_nnc_graph_exec_symbol_new(graph, layer_norm, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std }, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), "layer_norm"); | |||
2286 | outputs[0] = output; | |||
2287 | } | |||
2288 | ||||
2289 | static void _ccv_cnnp_layer_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
2290 | { | |||
2291 | ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super; | |||
2292 | if (self->scale.graph) | |||
2293 | initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale); | |||
2294 | if (self->bias.graph) | |||
2295 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
2296 | } | |||
2297 | ||||
2298 | static void _ccv_cnnp_layer_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
2299 | { | |||
2300 | ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super; | |||
2301 | if (self->scale.graph) | |||
2302 | add_to_array(parameters, self->scale, is_trainable); | |||
2303 | if (self->bias.graph) | |||
2304 | add_to_array(parameters, self->bias, is_trainable); | |||
2305 | } | |||
2306 | ||||
2307 | static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2308 | ||||
2309 | static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = { | |||
2310 | .build = _ccv_cnnp_layer_norm_build, | |||
2311 | .init_states = _ccv_cnnp_layer_norm_init_states, | |||
2312 | .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter, | |||
2313 | .copy = _ccv_cnnp_layer_norm_copy, | |||
2314 | }; | |||
2315 | ||||
2316 | ccv_cnnp_model_t* ccv_cnnp_layer_norm(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const int is_trainable, const char* const name) | |||
2317 | { | |||
2318 | ccv_cnnp_model_layer_norm_t* const model_layer_norm = (ccv_cnnp_model_layer_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_layer_norm_t)); | |||
2319 | model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa; | |||
2320 | model_layer_norm->super.input_size = 1; | |||
2321 | model_layer_norm->super.outputs = &model_layer_norm->output; | |||
2322 | model_layer_norm->super.output_size = 1; | |||
2323 | model_layer_norm->super.is_trainable = is_trainable; | |||
2324 | ccv_cnnp_model_copy_name(&model_layer_norm->super, name); | |||
2325 | model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2326 | model_layer_norm->scale.graph = 0; | |||
2327 | model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2328 | model_layer_norm->bias.graph = 0; | |||
2329 | model_layer_norm->params.lnorm.epsilon = epsilon; | |||
2330 | model_layer_norm->params.lnorm.count = axis_count; | |||
2331 | model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine; | |||
2332 | memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count); | |||
2333 | return (ccv_cnnp_model_t*)model_layer_norm; | |||
2334 | } | |||
2335 | ||||
2336 | static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2337 | { | |||
2338 | const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super; | |||
2339 | return ccv_cnnp_layer_norm(self->params.lnorm.epsilon, self->params.lnorm.axis, self->params.lnorm.count, self->params.lnorm.elementwise_affine, self->super.is_trainable, self->super.name); | |||
2340 | } | |||
2341 | ||||
2342 | // MARK - Group Norm Layer | |||
2343 | ||||
2344 | typedef struct { | |||
2345 | ccv_cnnp_model_t super; | |||
2346 | ccv_nnc_tensor_symbol_t output; | |||
2347 | ccv_nnc_tensor_symbol_t bias; | |||
2348 | ccv_nnc_tensor_symbol_t scale; | |||
2349 | ccv_nnc_cmd_param_t params; | |||
2350 | } ccv_cnnp_model_group_norm_t; | |||
2351 | ||||
2352 | static void _ccv_cnnp_group_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2353 | { | |||
2354 | PRINT(CCV_CLI_VERBOSE, "[cnnp_group_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_group_norm_build] -\n"); fflush(stdout); } } while (0); | |||
2355 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2355, __extension__ __PRETTY_FUNCTION__); })); | |||
2356 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2356, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2357 | ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super; | |||
2358 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2359 | ccv_nnc_tensor_param_t bias_params = params; | |||
2360 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
2361 | int i; | |||
2362 | for (i = 0; i < nd; i++) | |||
2363 | bias_params.dim[i] = 1; | |||
2364 | bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis]; | |||
2365 | if (self->params.gnorm.elementwise_affine) | |||
2366 | { | |||
2367 | // Both scale and bias are shared between if this model is reused. | |||
2368 | if (!self->scale.graph) | |||
2369 | self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale"); | |||
2370 | if (!self->bias.graph) | |||
2371 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
2372 | } | |||
2373 | const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0); | |||
2374 | ccv_nnc_tensor_param_t output_params[3]; | |||
2375 | if (self->params.gnorm.elementwise_affine) | |||
2376 | ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){ | |||
2377 | params, | |||
2378 | bias_params, | |||
2379 | bias_params, | |||
2380 | }, 3, ccv_nnc_no_hint, output_params, 3); | |||
2381 | else | |||
2382 | ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){ | |||
2383 | params, | |||
2384 | }, 1, ccv_nnc_no_hint, output_params, 3); | |||
2385 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
2386 | const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean"); | |||
2387 | const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std"); | |||
2388 | if (self->params.gnorm.elementwise_affine) | |||
2389 | ccv_nnc_graph_exec_symbol_new(graph, group_norm, TENSOR_SYMBOL_LIST(inputs[0], self->scale, self->bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->scale, self->bias}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std }, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), "group_norm"); | |||
2390 | else | |||
2391 | ccv_nnc_graph_exec_symbol_new(graph, group_norm, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std }, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), "group_norm"); | |||
2392 | outputs[0] = output; | |||
2393 | } | |||
2394 | ||||
2395 | static void _ccv_cnnp_group_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
2396 | { | |||
2397 | ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super; | |||
2398 | if (self->scale.graph) | |||
2399 | initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale); | |||
2400 | if (self->bias.graph) | |||
2401 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
2402 | } | |||
2403 | ||||
2404 | static void _ccv_cnnp_group_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
2405 | { | |||
2406 | ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super; | |||
2407 | if (self->scale.graph) | |||
2408 | add_to_array(parameters, self->scale, is_trainable); | |||
2409 | if (self->bias.graph) | |||
2410 | add_to_array(parameters, self->bias, is_trainable); | |||
2411 | } | |||
2412 | ||||
2413 | static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2414 | ||||
2415 | static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = { | |||
2416 | .build = _ccv_cnnp_group_norm_build, | |||
2417 | .init_states = _ccv_cnnp_group_norm_init_states, | |||
2418 | .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter, | |||
2419 | .copy = _ccv_cnnp_group_norm_copy, | |||
2420 | }; | |||
2421 | ||||
2422 | ccv_cnnp_model_t* ccv_cnnp_group_norm(const int group_axis, const int groups, const float epsilon, const int reduce_axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const int is_trainable, const char* const name) | |||
2423 | { | |||
2424 | ccv_cnnp_model_group_norm_t* const model_group_norm = (ccv_cnnp_model_group_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_group_norm_t)); | |||
2425 | model_group_norm->super.isa = &ccv_cnnp_group_norm_isa; | |||
2426 | model_group_norm->super.input_size = 1; | |||
2427 | model_group_norm->super.outputs = &model_group_norm->output; | |||
2428 | model_group_norm->super.output_size = 1; | |||
2429 | model_group_norm->super.is_trainable = is_trainable; | |||
2430 | ccv_cnnp_model_copy_name(&model_group_norm->super, name); | |||
2431 | model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2432 | model_group_norm->scale.graph = 0; | |||
2433 | model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2434 | model_group_norm->bias.graph = 0; | |||
2435 | model_group_norm->params.gnorm.group_axis = group_axis; | |||
2436 | model_group_norm->params.gnorm.groups = groups; | |||
2437 | model_group_norm->params.gnorm.epsilon = epsilon; | |||
2438 | model_group_norm->params.gnorm.reduce_count = axis_count; | |||
2439 | model_group_norm->params.gnorm.elementwise_affine = elementwise_affine; | |||
2440 | memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count); | |||
2441 | return (ccv_cnnp_model_t*)model_group_norm; | |||
2442 | } | |||
2443 | ||||
2444 | static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2445 | { | |||
2446 | const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super; | |||
2447 | return ccv_cnnp_group_norm(self->params.gnorm.group_axis, self->params.gnorm.groups, self->params.gnorm.epsilon, self->params.gnorm.reduce_axis, self->params.gnorm.reduce_count, self->params.gnorm.elementwise_affine, self->super.is_trainable, self->super.name); | |||
2448 | } | |||
2449 | ||||
2450 | // MARK - RMSNorm Layer | |||
2451 | ||||
2452 | typedef struct { | |||
2453 | ccv_cnnp_model_t super; | |||
2454 | ccv_nnc_tensor_symbol_t output; | |||
2455 | ccv_nnc_tensor_symbol_t scale; | |||
2456 | ccv_nnc_cmd_param_t params; | |||
2457 | } ccv_cnnp_model_rmsnorm_t; | |||
2458 | ||||
2459 | static void _ccv_cnnp_rmsnorm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2460 | { | |||
2461 | PRINT(CCV_CLI_VERBOSE, "[cnnp_rmsnorm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_rmsnorm_build] -\n"); fflush(stdout); } } while (0); | |||
2462 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2462, __extension__ __PRETTY_FUNCTION__); })); | |||
2463 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2463, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2464 | ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super; | |||
2465 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2466 | ccv_nnc_tensor_param_t scale_params = params; | |||
2467 | const int nd = ccv_nnc_tensor_nd(params.dim); | |||
2468 | int i; | |||
2469 | for (i = 0; i < nd; i++) | |||
2470 | scale_params.dim[i] = 1; | |||
2471 | for (i = 0; i < self->params.rmsnorm.count; i++) | |||
2472 | scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]]; | |||
2473 | // Both scale and bias are shared between if this model is reused. | |||
2474 | if (!self->scale.graph) | |||
2475 | self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale"); | |||
2476 | const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0); | |||
2477 | ccv_nnc_tensor_param_t output_params[2]; | |||
2478 | ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){ | |||
2479 | params, | |||
2480 | scale_params, | |||
2481 | }, 2, ccv_nnc_no_hint, output_params, 2); | |||
2482 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
2483 | const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std"); | |||
2484 | ccv_nnc_graph_exec_symbol_new(graph, rmsnorm, TENSOR_SYMBOL_LIST(inputs[0], self->scale)(const ccv_nnc_tensor_symbol_t []){inputs[0], self->scale} , (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_inv_std}, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 -1), "rmsnorm"); | |||
2485 | outputs[0] = output; | |||
2486 | } | |||
2487 | ||||
2488 | static void _ccv_cnnp_rmsnorm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
2489 | { | |||
2490 | ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super; | |||
2491 | if (self->scale.graph) | |||
2492 | initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale); | |||
2493 | } | |||
2494 | ||||
2495 | static void _ccv_cnnp_rmsnorm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
2496 | { | |||
2497 | ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super; | |||
2498 | if (self->scale.graph) | |||
2499 | add_to_array(parameters, self->scale, is_trainable); | |||
2500 | } | |||
2501 | ||||
2502 | static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2503 | ||||
2504 | static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = { | |||
2505 | .build = _ccv_cnnp_rmsnorm_build, | |||
2506 | .init_states = _ccv_cnnp_rmsnorm_init_states, | |||
2507 | .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter, | |||
2508 | .copy = _ccv_cnnp_rmsnorm_copy, | |||
2509 | }; | |||
2510 | ||||
2511 | ccv_cnnp_model_t* ccv_cnnp_rmsnorm(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int is_trainable, const char* const name) | |||
2512 | { | |||
2513 | ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t)); | |||
2514 | model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa; | |||
2515 | model_rmsnorm->super.input_size = 1; | |||
2516 | model_rmsnorm->super.outputs = &model_rmsnorm->output; | |||
2517 | model_rmsnorm->super.output_size = 1; | |||
2518 | model_rmsnorm->super.is_trainable = is_trainable; | |||
2519 | ccv_cnnp_model_copy_name(&model_rmsnorm->super, name); | |||
2520 | model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2521 | model_rmsnorm->scale.graph = 0; | |||
2522 | model_rmsnorm->params.rmsnorm.epsilon = epsilon; | |||
2523 | model_rmsnorm->params.rmsnorm.count = axis_count; | |||
2524 | memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count); | |||
2525 | return (ccv_cnnp_model_t*)model_rmsnorm; | |||
2526 | } | |||
2527 | ||||
2528 | static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2529 | { | |||
2530 | const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super; | |||
2531 | return ccv_cnnp_rmsnorm(self->params.rmsnorm.epsilon, self->params.rmsnorm.axis, self->params.rmsnorm.count, self->super.is_trainable, self->super.name); | |||
2532 | } | |||
2533 | ||||
2534 | // MARK - Batched Matrix Mul Layer | |||
2535 | ||||
2536 | typedef struct { | |||
2537 | ccv_cnnp_model_t super; | |||
2538 | ccv_nnc_tensor_symbol_t output; | |||
2539 | int transpose_a[2]; | |||
2540 | int transpose_b[2]; | |||
2541 | } ccv_cnnp_model_matmul_t; | |||
2542 | ||||
2543 | static void _ccv_cnnp_matmul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2544 | { | |||
2545 | PRINT(CCV_CLI_VERBOSE, "[cnnp_matmul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_matmul_build] -\n"); fflush(stdout); } } while (0); | |||
2546 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 2546, __extension__ __PRETTY_FUNCTION__); })); | |||
2547 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2547, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2548 | ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super; | |||
2549 | ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2550 | ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]); | |||
2551 | ccv_nnc_tensor_param_t output_params; | |||
2552 | const ccv_nnc_cmd_t matmul = CMD_GEMM_FORWARD(self->transpose_a, self->transpose_b)ccv_nnc_cmd(CCV_NNC_GEMM_FORWARD, 0, ((ccv_nnc_cmd_param_t){. size={.dim={1,1,1}},.blas={.a={1,1},.transpose_a={self->transpose_a [0],self->transpose_a[1]},.transpose_b={self->transpose_b [0],self->transpose_b[1]},}}), 0); | |||
2553 | ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){ | |||
2554 | a_params, | |||
2555 | b_params, | |||
2556 | }, 2, ccv_nnc_no_hint, &output_params, 1); | |||
2557 | const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2558 | ccv_nnc_graph_exec_symbol_new(graph, matmul, inputs, input_size, TENSOR_SYMBOL_LIST(matmul_output)(const ccv_nnc_tensor_symbol_t []){matmul_output}, (1 +1 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1 ), "matmul"); | |||
2559 | outputs[0] = matmul_output; | |||
2560 | } | |||
2561 | ||||
2562 | static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2563 | ||||
2564 | static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = { | |||
2565 | .build = _ccv_cnnp_matmul_build, | |||
2566 | .copy = _ccv_cnnp_matmul_copy, | |||
2567 | }; | |||
2568 | ||||
2569 | ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const char* const name) | |||
2570 | { | |||
2571 | ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t)); | |||
2572 | model_matmul->super.isa = &ccv_cnnp_matmul_isa; | |||
2573 | model_matmul->super.input_size = 2; | |||
2574 | model_matmul->super.outputs = &model_matmul->output; | |||
2575 | model_matmul->super.output_size = 1; | |||
2576 | model_matmul->transpose_a[0] = transpose_a[0]; | |||
2577 | model_matmul->transpose_a[1] = transpose_a[1]; | |||
2578 | model_matmul->transpose_b[0] = transpose_b[0]; | |||
2579 | model_matmul->transpose_b[1] = transpose_b[1]; | |||
2580 | ccv_cnnp_model_copy_name(&model_matmul->super, name); | |||
2581 | return (ccv_cnnp_model_t*)model_matmul; | |||
2582 | } | |||
2583 | ||||
2584 | static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2585 | { | |||
2586 | const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super; | |||
2587 | return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->super.name); | |||
2588 | } | |||
2589 | ||||
2590 | // MARK - Dropout Layer | |||
2591 | ||||
2592 | typedef struct { | |||
2593 | ccv_cnnp_model_t super; | |||
2594 | ccv_nnc_tensor_symbol_t output; | |||
2595 | ccv_nnc_graph_exec_symbol_t dropout; | |||
2596 | float p; | |||
2597 | int entirety; | |||
2598 | } ccv_cnnp_model_dropout_t; | |||
2599 | ||||
2600 | static void _ccv_cnnp_dropout_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2601 | { | |||
2602 | PRINT(CCV_CLI_VERBOSE, "[cnnp_dropout_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_dropout_build] -\n"); fflush(stdout); } } while (0); | |||
2603 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2603, __extension__ __PRETTY_FUNCTION__); })); | |||
2604 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2604, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2605 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2606 | ccv_nnc_tensor_param_t output_params[2]; | |||
2607 | ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super; | |||
2608 | const ccv_nnc_cmd_t dropout = CMD_DROPOUT_FORWARD(self->p, self->entirety)ccv_nnc_cmd(CCV_NNC_DROPOUT_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.dropout={.p=self->p,.entirety=self ->entirety}}), 0); | |||
2609 | ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){ | |||
2610 | params, | |||
2611 | }, 1, ccv_nnc_no_hint, output_params, 2); | |||
2612 | const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
2613 | const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask"); | |||
2614 | self->dropout = ccv_nnc_graph_exec_symbol_new(graph, dropout, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(dropout_output, mask)(const ccv_nnc_tensor_symbol_t []){dropout_output, mask}, (1 + 1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dropout"); | |||
2615 | outputs[0] = dropout_output; | |||
2616 | } | |||
2617 | ||||
2618 | static void _ccv_cnnp_dropout_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context) | |||
2619 | { | |||
2620 | ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super; | |||
2621 | if (self->dropout.graph) | |||
2622 | { | |||
2623 | if (is_test) | |||
2624 | // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip. | |||
2625 | updater(context, self->dropout, CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto , 0), ccv_nnc_no_hint); | |||
2626 | else | |||
2627 | updater(context, self->dropout, CMD_DROPOUT_FORWARD(self->p, self->entirety)ccv_nnc_cmd(CCV_NNC_DROPOUT_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.dropout={.p=self->p,.entirety=self ->entirety}}), 0), ccv_nnc_no_hint); | |||
2628 | } | |||
2629 | } | |||
2630 | ||||
2631 | static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2632 | ||||
2633 | static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = { | |||
2634 | .build = _ccv_cnnp_dropout_build, | |||
2635 | .set_is_test = _ccv_cnnp_dropout_set_is_test, | |||
2636 | .copy = _ccv_cnnp_dropout_copy, | |||
2637 | }; | |||
2638 | ||||
2639 | ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name) | |||
2640 | { | |||
2641 | ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t)); | |||
2642 | model_dropout->super.isa = &ccv_cnnp_dropout_isa; | |||
2643 | model_dropout->super.input_size = 1; | |||
2644 | model_dropout->super.outputs = &model_dropout->output; | |||
2645 | model_dropout->super.output_size = 1; | |||
2646 | model_dropout->p = p; | |||
2647 | model_dropout->entirety = entirety; | |||
2648 | ccv_cnnp_model_copy_name(&model_dropout->super, name); | |||
2649 | return (ccv_cnnp_model_t*)model_dropout; | |||
2650 | } | |||
2651 | ||||
2652 | static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2653 | { | |||
2654 | const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super; | |||
2655 | return ccv_cnnp_dropout(self->p, self->entirety, self->super.name); | |||
2656 | } | |||
2657 | ||||
2658 | // MARK - Masked Fill Layer | |||
2659 | ||||
2660 | typedef struct { | |||
2661 | ccv_cnnp_model_t super; | |||
2662 | ccv_nnc_tensor_symbol_t output; | |||
2663 | float eq; | |||
2664 | float fill; | |||
2665 | } ccv_cnnp_model_masked_fill_t; | |||
2666 | ||||
2667 | static void _ccv_cnnp_masked_fill_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2668 | { | |||
2669 | PRINT(CCV_CLI_VERBOSE, "[cnnp_masked_fill_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_masked_fill_build] -\n"); fflush(stdout); } } while (0); | |||
2670 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 2670, __extension__ __PRETTY_FUNCTION__); })); | |||
2671 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2671, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2672 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2673 | ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super; | |||
2674 | const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
2675 | ccv_nnc_graph_exec_symbol_new(graph, CMD_MASKED_FILL_FORWARD(self->eq, self->fill)ccv_nnc_cmd(CCV_NNC_MASKED_FILL_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={self->eq, self->fill} }}, 0), TENSOR_SYMBOL_LIST(inputs[0], inputs[1])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1]}, (1 + 1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(masked_fill_output)(const ccv_nnc_tensor_symbol_t []){masked_fill_output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 -1), "masked_fill"); | |||
2676 | outputs[0] = masked_fill_output; | |||
2677 | } | |||
2678 | ||||
2679 | static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2680 | ||||
2681 | static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = { | |||
2682 | .build = _ccv_cnnp_masked_fill_build, | |||
2683 | .copy = _ccv_cnnp_masked_fill_copy, | |||
2684 | }; | |||
2685 | ||||
2686 | ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name) | |||
2687 | { | |||
2688 | ccv_cnnp_model_masked_fill_t* const model_masked_fill = (ccv_cnnp_model_masked_fill_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_masked_fill_t)); | |||
2689 | model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa; | |||
2690 | model_masked_fill->super.input_size = 2; | |||
2691 | model_masked_fill->super.outputs = &model_masked_fill->output; | |||
2692 | model_masked_fill->super.output_size = 1; | |||
2693 | model_masked_fill->eq = eq; | |||
2694 | model_masked_fill->fill = fill; | |||
2695 | ccv_cnnp_model_copy_name(&model_masked_fill->super, name); | |||
2696 | return (ccv_cnnp_model_t*)model_masked_fill; | |||
2697 | } | |||
2698 | ||||
2699 | static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2700 | { | |||
2701 | const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super; | |||
2702 | return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name); | |||
2703 | } | |||
2704 | ||||
2705 | // MARK - Index Select Layer | |||
2706 | ||||
2707 | typedef struct { | |||
2708 | ccv_cnnp_model_t super; | |||
2709 | ccv_nnc_tensor_symbol_t output; | |||
2710 | } ccv_cnnp_model_index_select_t; | |||
2711 | ||||
2712 | static void _ccv_cnnp_index_select_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2713 | { | |||
2714 | PRINT(CCV_CLI_VERBOSE, "[cnnp_index_select_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_index_select_build] -\n"); fflush(stdout); } } while (0); | |||
2715 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 2715, __extension__ __PRETTY_FUNCTION__); })); | |||
2716 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2716, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2717 | const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2718 | const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]); | |||
2719 | ccv_nnc_tensor_param_t output_params; | |||
2720 | const ccv_nnc_cmd_t index_select = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto , 0); | |||
2721 | ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){ | |||
2722 | vocab_params, | |||
2723 | index_params, | |||
2724 | }, 2, ccv_nnc_no_hint, &output_params, 1); | |||
2725 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2726 | ccv_nnc_graph_exec_symbol_new(graph, index_select, TENSOR_SYMBOL_LIST(inputs[0], inputs[1])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1]}, (1 + 1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "index_select"); | |||
2727 | outputs[0] = output; | |||
2728 | } | |||
2729 | ||||
2730 | static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2731 | ||||
2732 | static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = { | |||
2733 | .build = _ccv_cnnp_index_select_build, | |||
2734 | .copy = _ccv_cnnp_index_select_copy, | |||
2735 | }; | |||
2736 | ||||
2737 | ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name) | |||
2738 | { | |||
2739 | ccv_cnnp_model_index_select_t* const model_index_select = (ccv_cnnp_model_index_select_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_index_select_t)); | |||
2740 | model_index_select->super.isa = &ccv_cnnp_index_select_isa; | |||
2741 | model_index_select->super.input_size = 2; | |||
2742 | model_index_select->super.outputs = &model_index_select->output; | |||
2743 | model_index_select->super.output_size = 1; | |||
2744 | ccv_cnnp_model_copy_name(&model_index_select->super, name); | |||
2745 | return (ccv_cnnp_model_t*)model_index_select; | |||
2746 | } | |||
2747 | ||||
2748 | static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2749 | { | |||
2750 | ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super; | |||
2751 | return ccv_cnnp_index_select(self->super.name); | |||
2752 | } | |||
2753 | ||||
2754 | // MARK - Embedding Layer | |||
2755 | ||||
2756 | typedef struct { | |||
2757 | ccv_cnnp_model_t super; | |||
2758 | ccv_nnc_tensor_symbol_t output; | |||
2759 | ccv_nnc_tensor_symbol_t vocab; | |||
2760 | int datatype; | |||
2761 | int vocab_size; | |||
2762 | int embed_size; | |||
2763 | } ccv_cnnp_model_embedding_t; | |||
2764 | ||||
2765 | static void _ccv_cnnp_embedding_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2766 | { | |||
2767 | ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super; | |||
2768 | PRINT(CCV_CLI_VERBOSE, "[cnnp_embedding_build] vocab_size: %d, embed_size: %d\n", self->vocab_size, self->embed_size)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_embedding_build] vocab_size: %d, embed_size: %d\n" , self->vocab_size, self->embed_size); fflush(stdout); } } while (0); | |||
2769 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2769, __extension__ __PRETTY_FUNCTION__); })); | |||
2770 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2770, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2771 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2772 | ccv_nnc_tensor_param_t vocab_params = params; | |||
2773 | memset(vocab_params.dim, 0, sizeof(vocab_params.dim)); | |||
2774 | vocab_params.datatype = self->datatype; | |||
2775 | vocab_params.dim[0] = self->vocab_size; | |||
2776 | vocab_params.dim[1] = self->embed_size; | |||
2777 | if (!self->vocab.graph) | |||
2778 | self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab"); | |||
2779 | assert(self->vocab.graph == graph)((void) sizeof ((self->vocab.graph == graph) ? 1 : 0), __extension__ ({ if (self->vocab.graph == graph) ; else __assert_fail ( "self->vocab.graph == graph", "ccv_cnnp_model_addons.c", 2779 , __extension__ __PRETTY_FUNCTION__); })); | |||
2780 | ccv_nnc_tensor_param_t output_params; | |||
2781 | const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto , 0); | |||
2782 | ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){ | |||
2783 | vocab_params, | |||
2784 | params, | |||
2785 | }, 2, ccv_nnc_no_hint, &output_params, 1); | |||
2786 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2787 | ccv_nnc_graph_exec_symbol_new(graph, embedding, TENSOR_SYMBOL_LIST(self->vocab, inputs[0])(const ccv_nnc_tensor_symbol_t []){self->vocab, inputs[0]} , (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "embedding"); | |||
2788 | outputs[0] = output; | |||
2789 | } | |||
2790 | ||||
2791 | static void _ccv_cnnp_embedding_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
2792 | { | |||
2793 | ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super; | |||
2794 | const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size); | |||
2795 | const float bound = sqrtf(3) * std; | |||
2796 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->vocab); | |||
2797 | } | |||
2798 | ||||
2799 | static void _ccv_cnnp_embedding_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
2800 | { | |||
2801 | ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super; | |||
2802 | add_to_array(parameters, self->vocab, is_trainable); | |||
2803 | } | |||
2804 | ||||
2805 | static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2806 | ||||
2807 | static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = { | |||
2808 | .build = _ccv_cnnp_embedding_build, | |||
2809 | .init_states = _ccv_cnnp_embedding_init_states, | |||
2810 | .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter, | |||
2811 | .copy = _ccv_cnnp_embedding_copy, | |||
2812 | }; | |||
2813 | ||||
2814 | ccv_cnnp_model_t* ccv_cnnp_embedding(const int datatype, const int vocab_size, const int embed_size, const int is_trainable, const char* const name) | |||
2815 | { | |||
2816 | ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t)); | |||
2817 | model_embedding->super.isa = &ccv_cnnp_embedding_isa; | |||
2818 | model_embedding->super.input_size = 1; | |||
2819 | model_embedding->super.outputs = &model_embedding->output; | |||
2820 | model_embedding->super.output_size = 1; | |||
2821 | model_embedding->super.is_trainable = is_trainable; | |||
2822 | ccv_cnnp_model_copy_name(&model_embedding->super, name); | |||
2823 | model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
2824 | model_embedding->vocab.graph = 0; | |||
2825 | assert(datatype == CCV_32F || datatype == CCV_16F)((void) sizeof ((datatype == CCV_32F || datatype == CCV_16F) ? 1 : 0), __extension__ ({ if (datatype == CCV_32F || datatype == CCV_16F) ; else __assert_fail ("datatype == CCV_32F || datatype == CCV_16F" , "ccv_cnnp_model_addons.c", 2825, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2826 | model_embedding->datatype = datatype; | |||
2827 | assert(vocab_size > 0)((void) sizeof ((vocab_size > 0) ? 1 : 0), __extension__ ( { if (vocab_size > 0) ; else __assert_fail ("vocab_size > 0" , "ccv_cnnp_model_addons.c", 2827, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2828 | model_embedding->vocab_size = vocab_size; | |||
2829 | assert(embed_size > 0)((void) sizeof ((embed_size > 0) ? 1 : 0), __extension__ ( { if (embed_size > 0) ; else __assert_fail ("embed_size > 0" , "ccv_cnnp_model_addons.c", 2829, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2830 | model_embedding->embed_size = embed_size; | |||
2831 | return (ccv_cnnp_model_t*)model_embedding; | |||
2832 | } | |||
2833 | ||||
2834 | static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2835 | { | |||
2836 | ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super; | |||
2837 | return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name); | |||
2838 | } | |||
2839 | ||||
2840 | // MARK - Pool Layers | |||
2841 | ||||
2842 | typedef struct { | |||
2843 | ccv_cnnp_model_t super; | |||
2844 | ccv_nnc_tensor_symbol_t output; | |||
2845 | int type; | |||
2846 | float width_scale; | |||
2847 | float height_scale; | |||
2848 | int align_corners; | |||
2849 | } ccv_cnnp_model_upsample_t; | |||
2850 | ||||
2851 | static void _ccv_cnnp_upsample_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2852 | { | |||
2853 | PRINT(CCV_CLI_VERBOSE, "[cnnp_upsample_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_upsample_build] -\n"); fflush(stdout); } } while (0); | |||
2854 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2854, __extension__ __PRETTY_FUNCTION__); })); | |||
2855 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2855, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2856 | ccv_cnnp_model_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super; | |||
2857 | const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2858 | ccv_nnc_cmd_t cmd = CMD_UPSAMPLE_FORWARD(self->type, self->width_scale, self->height_scale, self->align_corners)ccv_nnc_cmd(CCV_NNC_UPSAMPLE_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.upsample={.type=self->type,.width_scale =self->width_scale,.height_scale=self->height_scale,.align_corners =self->align_corners}}), 0); | |||
2859 | ccv_nnc_tensor_param_t output_params; | |||
2860 | ccv_nnc_hint_tensor_auto(cmd, ¶ms, 1, ccv_nnc_no_hint, &output_params, 1); | |||
2861 | const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2862 | ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "upsample"); | |||
2863 | outputs[0] = output; | |||
2864 | } | |||
2865 | ||||
2866 | static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
2867 | ||||
2868 | static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = { | |||
2869 | .build = _ccv_cnnp_upsample_build, | |||
2870 | .copy = _ccv_cnnp_upsample_copy, | |||
2871 | }; | |||
2872 | ||||
2873 | ccv_cnnp_model_t* ccv_cnnp_upsample(const int type, const float width_scale, const float height_scale, const int align_corners, const char* const name) | |||
2874 | { | |||
2875 | ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t)); | |||
2876 | model_upsample->super.isa = &ccv_cnnp_upsample_isa; | |||
2877 | model_upsample->super.input_size = 1; | |||
2878 | model_upsample->super.outputs = &model_upsample->output; | |||
2879 | model_upsample->super.output_size = 1; | |||
2880 | ccv_cnnp_model_copy_name(&model_upsample->super, name); | |||
2881 | assert(type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR)((void) sizeof ((type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR ) ? 1 : 0), __extension__ ({ if (type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR) ; else __assert_fail ( "type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR" , "ccv_cnnp_model_addons.c", 2881, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2882 | model_upsample->type = type; | |||
2883 | model_upsample->width_scale = width_scale; | |||
2884 | model_upsample->height_scale = height_scale; | |||
2885 | model_upsample->align_corners = align_corners; | |||
2886 | return (ccv_cnnp_model_t*)model_upsample; | |||
2887 | } | |||
2888 | ||||
2889 | static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2890 | { | |||
2891 | const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super; | |||
2892 | return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name); | |||
2893 | } | |||
2894 | ||||
2895 | // MARK - Reduce Sum Layer | |||
2896 | ||||
2897 | typedef struct { | |||
2898 | ccv_cnnp_model_t super; | |||
2899 | int axis[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
2900 | int count; | |||
2901 | ccv_nnc_tensor_symbol_t output; | |||
2902 | } ccv_cnnp_model_reduce_sum_t; | |||
2903 | ||||
2904 | static void _ccv_cnnp_reduce_sum_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2905 | { | |||
2906 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_sum_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reduce_sum_build] -\n"); fflush(stdout); } } while (0); | |||
2907 | const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super; | |||
2908 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2908, __extension__ __PRETTY_FUNCTION__); })); | |||
2909 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2909, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2910 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2911 | ccv_nnc_tensor_param_t output_params; | |||
2912 | ccv_nnc_cmd_t reduce_sum = CMD_REDUCE_SUM_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_SUM_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
2913 | int i; | |||
2914 | for (i = 0; i < self->count; i++) | |||
2915 | reduce_sum.info.reduce.axis[i] = self->axis[i]; | |||
2916 | reduce_sum.info.reduce.count = self->count; | |||
2917 | ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
2918 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2919 | ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum"); | |||
2920 | } | |||
2921 | ||||
2922 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
2923 | ||||
2924 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = { | |||
2925 | .build = _ccv_cnnp_reduce_sum_build, | |||
2926 | .copy = _ccv_cnnp_reduce_sum_copy, | |||
2927 | }; | |||
2928 | ||||
2929 | ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name) | |||
2930 | { | |||
2931 | ccv_cnnp_model_reduce_sum_t* const model_reduce_sum = (ccv_cnnp_model_reduce_sum_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_sum_t)); | |||
2932 | model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa; | |||
2933 | model_reduce_sum->super.input_size = 1; | |||
2934 | model_reduce_sum->super.outputs = &model_reduce_sum->output; | |||
2935 | model_reduce_sum->super.output_size = 1; | |||
2936 | ccv_cnnp_model_copy_name(&model_reduce_sum->super, name); | |||
2937 | assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__ ({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC" , "ccv_cnnp_model_addons.c", 2937, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2938 | int i; | |||
2939 | for (i = 0; i < axis_count; i++) | |||
2940 | model_reduce_sum->axis[i] = axis[i]; | |||
2941 | model_reduce_sum->count = axis_count; | |||
2942 | return (ccv_cnnp_model_t*)model_reduce_sum; | |||
2943 | } | |||
2944 | ||||
2945 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
2946 | { | |||
2947 | const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super; | |||
2948 | return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name); | |||
2949 | } | |||
2950 | ||||
2951 | // MARK - Reduce Mean Layer | |||
2952 | ||||
2953 | typedef struct { | |||
2954 | ccv_cnnp_model_t super; | |||
2955 | int axis[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
2956 | int count; | |||
2957 | ccv_nnc_tensor_symbol_t output; | |||
2958 | } ccv_cnnp_model_reduce_mean_t; | |||
2959 | ||||
2960 | static void _ccv_cnnp_reduce_mean_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
2961 | { | |||
2962 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_mean_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reduce_mean_build] -\n"); fflush(stdout); } } while (0); | |||
2963 | const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super; | |||
2964 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 2964, __extension__ __PRETTY_FUNCTION__); })); | |||
2965 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 2965, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2966 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
2967 | ccv_nnc_tensor_param_t output_params; | |||
2968 | ccv_nnc_cmd_t reduce_mean = CMD_REDUCE_MEAN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MEAN_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
2969 | int i; | |||
2970 | for (i = 0; i < self->count; i++) | |||
2971 | reduce_mean.info.reduce.axis[i] = self->axis[i]; | |||
2972 | reduce_mean.info.reduce.count = self->count; | |||
2973 | ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
2974 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
2975 | ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean"); | |||
2976 | } | |||
2977 | ||||
2978 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
2979 | ||||
2980 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = { | |||
2981 | .build = _ccv_cnnp_reduce_mean_build, | |||
2982 | .copy = _ccv_cnnp_reduce_mean_copy, | |||
2983 | }; | |||
2984 | ||||
2985 | ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name) | |||
2986 | { | |||
2987 | ccv_cnnp_model_reduce_mean_t* const model_reduce_mean = (ccv_cnnp_model_reduce_mean_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_mean_t)); | |||
2988 | model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa; | |||
2989 | model_reduce_mean->super.input_size = 1; | |||
2990 | model_reduce_mean->super.outputs = &model_reduce_mean->output; | |||
2991 | model_reduce_mean->super.output_size = 1; | |||
2992 | ccv_cnnp_model_copy_name(&model_reduce_mean->super, name); | |||
2993 | assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__ ({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC" , "ccv_cnnp_model_addons.c", 2993, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2994 | int i; | |||
2995 | for (i = 0; i < axis_count; i++) | |||
2996 | model_reduce_mean->axis[i] = axis[i]; | |||
2997 | model_reduce_mean->count = axis_count; | |||
2998 | return (ccv_cnnp_model_t*)model_reduce_mean; | |||
2999 | } | |||
3000 | ||||
3001 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3002 | { | |||
3003 | const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super; | |||
3004 | return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name); | |||
3005 | } | |||
3006 | ||||
3007 | // MARK - Reduce Max Layer | |||
3008 | ||||
3009 | typedef struct { | |||
3010 | ccv_cnnp_model_t super; | |||
3011 | int axis[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
3012 | int count; | |||
3013 | ccv_nnc_tensor_symbol_t output; | |||
3014 | } ccv_cnnp_model_reduce_max_t; | |||
3015 | ||||
3016 | static void _ccv_cnnp_reduce_max_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3017 | { | |||
3018 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_max_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reduce_max_build] -\n"); fflush(stdout); } } while (0); | |||
3019 | const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super; | |||
3020 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3020, __extension__ __PRETTY_FUNCTION__); })); | |||
3021 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3021, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3022 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3023 | ccv_nnc_tensor_param_t output_params; | |||
3024 | ccv_nnc_cmd_t reduce_max = CMD_REDUCE_MAX_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MAX_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
3025 | int i; | |||
3026 | for (i = 0; i < self->count; i++) | |||
3027 | reduce_max.info.reduce.axis[i] = self->axis[i]; | |||
3028 | reduce_max.info.reduce.count = self->count; | |||
3029 | ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
3030 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3031 | ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max"); | |||
3032 | } | |||
3033 | ||||
3034 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3035 | ||||
3036 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = { | |||
3037 | .build = _ccv_cnnp_reduce_max_build, | |||
3038 | .copy = _ccv_cnnp_reduce_max_copy, | |||
3039 | }; | |||
3040 | ||||
3041 | ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name) | |||
3042 | { | |||
3043 | ccv_cnnp_model_reduce_max_t* const model_reduce_max = (ccv_cnnp_model_reduce_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_max_t)); | |||
3044 | model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa; | |||
3045 | model_reduce_max->super.input_size = 1; | |||
3046 | model_reduce_max->super.outputs = &model_reduce_max->output; | |||
3047 | model_reduce_max->super.output_size = 1; | |||
3048 | ccv_cnnp_model_copy_name(&model_reduce_max->super, name); | |||
3049 | assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__ ({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC" , "ccv_cnnp_model_addons.c", 3049, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3050 | int i; | |||
3051 | for (i = 0; i < axis_count; i++) | |||
3052 | model_reduce_max->axis[i] = axis[i]; | |||
3053 | model_reduce_max->count = axis_count; | |||
3054 | return (ccv_cnnp_model_t*)model_reduce_max; | |||
3055 | } | |||
3056 | ||||
3057 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3058 | { | |||
3059 | const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super; | |||
3060 | return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name); | |||
3061 | } | |||
3062 | ||||
3063 | // MARK - Reduce Min Layer | |||
3064 | ||||
3065 | typedef struct { | |||
3066 | ccv_cnnp_model_t super; | |||
3067 | int axis[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
3068 | int count; | |||
3069 | ccv_nnc_tensor_symbol_t output; | |||
3070 | } ccv_cnnp_model_reduce_min_t; | |||
3071 | ||||
3072 | static void _ccv_cnnp_reduce_min_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3073 | { | |||
3074 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_min_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reduce_min_build] -\n"); fflush(stdout); } } while (0); | |||
3075 | const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super; | |||
3076 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3076, __extension__ __PRETTY_FUNCTION__); })); | |||
3077 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3077, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3078 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3079 | ccv_nnc_tensor_param_t output_params; | |||
3080 | ccv_nnc_cmd_t reduce_min = CMD_REDUCE_MIN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MIN_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
3081 | int i; | |||
3082 | for (i = 0; i < self->count; i++) | |||
3083 | reduce_min.info.reduce.axis[i] = self->axis[i]; | |||
3084 | reduce_min.info.reduce.count = self->count; | |||
3085 | ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
3086 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3087 | ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min"); | |||
3088 | } | |||
3089 | ||||
3090 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3091 | ||||
3092 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = { | |||
3093 | .build = _ccv_cnnp_reduce_min_build, | |||
3094 | .copy = _ccv_cnnp_reduce_min_copy, | |||
3095 | }; | |||
3096 | ||||
3097 | ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name) | |||
3098 | { | |||
3099 | ccv_cnnp_model_reduce_min_t* const model_reduce_min = (ccv_cnnp_model_reduce_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_min_t)); | |||
3100 | model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa; | |||
3101 | model_reduce_min->super.input_size = 1; | |||
3102 | model_reduce_min->super.outputs = &model_reduce_min->output; | |||
3103 | model_reduce_min->super.output_size = 1; | |||
3104 | ccv_cnnp_model_copy_name(&model_reduce_min->super, name); | |||
3105 | assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__ ({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC" , "ccv_cnnp_model_addons.c", 3105, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3106 | int i; | |||
3107 | for (i = 0; i < axis_count; i++) | |||
3108 | model_reduce_min->axis[i] = axis[i]; | |||
3109 | model_reduce_min->count = axis_count; | |||
3110 | return (ccv_cnnp_model_t*)model_reduce_min; | |||
3111 | } | |||
3112 | ||||
3113 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3114 | { | |||
3115 | const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super; | |||
3116 | return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name); | |||
3117 | } | |||
3118 | ||||
3119 | // MARK - Reduce Norm2 Layer | |||
3120 | ||||
3121 | typedef struct { | |||
3122 | ccv_cnnp_model_t super; | |||
3123 | int axis[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
3124 | int count; | |||
3125 | ccv_nnc_tensor_symbol_t output; | |||
3126 | } ccv_cnnp_model_reduce_norm2_t; | |||
3127 | ||||
3128 | static void _ccv_cnnp_reduce_norm2_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3129 | { | |||
3130 | const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super; | |||
3131 | PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_norm2_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_reduce_norm2_build] -\n"); fflush(stdout); } } while (0); | |||
3132 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3132, __extension__ __PRETTY_FUNCTION__); })); | |||
3133 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3133, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3134 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3135 | ccv_nnc_tensor_param_t output_params; | |||
3136 | ccv_nnc_cmd_t reduce_norm2 = CMD_REDUCE_NORM2_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_NORM2_FORWARD, 0, ((ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}} ), 0); | |||
3137 | int i; | |||
3138 | for (i = 0; i < self->count; i++) | |||
3139 | reduce_norm2.info.reduce.axis[i] = self->axis[i]; | |||
3140 | reduce_norm2.info.reduce.count = self->count; | |||
3141 | ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
3142 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3143 | ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2"); | |||
3144 | } | |||
3145 | ||||
3146 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3147 | ||||
3148 | static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = { | |||
3149 | .build = _ccv_cnnp_reduce_norm2_build, | |||
3150 | .copy = _ccv_cnnp_reduce_norm2_copy, | |||
3151 | }; | |||
3152 | ||||
3153 | ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name) | |||
3154 | { | |||
3155 | ccv_cnnp_model_reduce_norm2_t* const model_reduce_norm2 = (ccv_cnnp_model_reduce_norm2_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_norm2_t)); | |||
3156 | model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa; | |||
3157 | model_reduce_norm2->super.input_size = 1; | |||
3158 | model_reduce_norm2->super.outputs = &model_reduce_norm2->output; | |||
3159 | model_reduce_norm2->super.output_size = 1; | |||
3160 | ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name); | |||
3161 | assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__ ({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC" , "ccv_cnnp_model_addons.c", 3161, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3162 | int i; | |||
3163 | for (i = 0; i < axis_count; i++) | |||
3164 | model_reduce_norm2->axis[i] = axis[i]; | |||
3165 | model_reduce_norm2->count = axis_count; | |||
3166 | return (ccv_cnnp_model_t*)model_reduce_norm2; | |||
3167 | } | |||
3168 | ||||
3169 | static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3170 | { | |||
3171 | const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super; | |||
3172 | return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name); | |||
3173 | } | |||
3174 | ||||
3175 | // MARK - Argmax Layer | |||
3176 | ||||
3177 | typedef struct { | |||
3178 | ccv_cnnp_model_t super; | |||
3179 | int axis; | |||
3180 | ccv_nnc_tensor_symbol_t output; | |||
3181 | } ccv_cnnp_model_argmax_t; | |||
3182 | ||||
3183 | static void _ccv_cnnp_argmax_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3184 | { | |||
3185 | const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super; | |||
3186 | PRINT(CCV_CLI_VERBOSE, "[cnnp_argmax_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_argmax_build] -\n"); fflush(stdout); } } while (0); | |||
3187 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3187, __extension__ __PRETTY_FUNCTION__); })); | |||
3188 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3188, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3189 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3190 | ccv_nnc_tensor_param_t output_params; | |||
3191 | ccv_nnc_cmd_t argmax = CMD_ARGMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_ARGMAX_FORWARD, 0, ((ccv_nnc_cmd_param_t) {.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}) , 0); | |||
3192 | argmax.info.reduce.axis[0] = self->axis; | |||
3193 | argmax.info.reduce.count = 1; | |||
3194 | ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
3195 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3196 | ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax"); | |||
3197 | } | |||
3198 | ||||
3199 | static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3200 | ||||
3201 | static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = { | |||
3202 | .build = _ccv_cnnp_argmax_build, | |||
3203 | .copy = _ccv_cnnp_argmax_copy, | |||
3204 | }; | |||
3205 | ||||
3206 | ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name) | |||
3207 | { | |||
3208 | ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t)); | |||
3209 | model_argmax->super.isa = &ccv_cnnp_argmax_isa; | |||
3210 | model_argmax->super.input_size = 1; | |||
3211 | model_argmax->super.outputs = &model_argmax->output; | |||
3212 | model_argmax->super.output_size = 1; | |||
3213 | ccv_cnnp_model_copy_name(&model_argmax->super, name); | |||
3214 | model_argmax->axis = axis; | |||
3215 | return (ccv_cnnp_model_t*)model_argmax; | |||
3216 | } | |||
3217 | ||||
3218 | static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3219 | { | |||
3220 | const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super; | |||
3221 | return ccv_cnnp_argmax(self->axis, self->super.name); | |||
3222 | } | |||
3223 | ||||
3224 | // MARK - Argmin Layer | |||
3225 | ||||
3226 | typedef struct { | |||
3227 | ccv_cnnp_model_t super; | |||
3228 | int axis; | |||
3229 | ccv_nnc_tensor_symbol_t output; | |||
3230 | } ccv_cnnp_model_argmin_t; | |||
3231 | ||||
3232 | static void _ccv_cnnp_argmin_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3233 | { | |||
3234 | const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super; | |||
3235 | PRINT(CCV_CLI_VERBOSE, "[cnnp_argmin_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_argmin_build] -\n"); fflush(stdout); } } while (0); | |||
3236 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3236, __extension__ __PRETTY_FUNCTION__); })); | |||
3237 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3237, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3238 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3239 | ccv_nnc_tensor_param_t output_params; | |||
3240 | ccv_nnc_cmd_t argmin = CMD_ARGMIN_FORWARD()ccv_nnc_cmd(CCV_NNC_ARGMIN_FORWARD, 0, ((ccv_nnc_cmd_param_t) {.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}) , 0); | |||
3241 | argmin.info.reduce.axis[0] = self->axis; | |||
3242 | argmin.info.reduce.count = 1; | |||
3243 | ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1); | |||
3244 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3245 | ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin"); | |||
3246 | } | |||
3247 | ||||
3248 | static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3249 | ||||
3250 | static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = { | |||
3251 | .build = _ccv_cnnp_argmin_build, | |||
3252 | .copy = _ccv_cnnp_argmin_copy, | |||
3253 | }; | |||
3254 | ||||
3255 | ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name) | |||
3256 | { | |||
3257 | ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t)); | |||
3258 | model_argmin->super.isa = &ccv_cnnp_argmin_isa; | |||
3259 | model_argmin->super.input_size = 1; | |||
3260 | model_argmin->super.outputs = &model_argmin->output; | |||
3261 | model_argmin->super.output_size = 1; | |||
3262 | ccv_cnnp_model_copy_name(&model_argmin->super, name); | |||
3263 | model_argmin->axis = axis; | |||
3264 | return (ccv_cnnp_model_t*)model_argmin; | |||
3265 | } | |||
3266 | ||||
3267 | static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3268 | { | |||
3269 | const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super; | |||
3270 | return ccv_cnnp_argmin(self->axis, self->super.name); | |||
3271 | } | |||
3272 | ||||
3273 | // MARK - Min Layer | |||
3274 | ||||
3275 | typedef struct { | |||
3276 | ccv_cnnp_model_t super; | |||
3277 | ccv_nnc_tensor_symbol_t output; | |||
3278 | } ccv_cnnp_model_min_t; | |||
3279 | ||||
3280 | static void _ccv_cnnp_min_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3281 | { | |||
3282 | PRINT(CCV_CLI_VERBOSE, "[cnnp_min_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_min_build] -\n"); fflush(stdout); } } while ( 0); | |||
3283 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 3283, __extension__ __PRETTY_FUNCTION__); })); | |||
3284 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3284, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3285 | ccv_nnc_tensor_param_t input_params[2]; | |||
3286 | int i; | |||
3287 | for (i = 0; i < 2; i++) | |||
3288 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
3289 | ccv_nnc_tensor_param_t output_params; | |||
3290 | const ccv_nnc_cmd_t min = CMD_MIN_FORWARD()ccv_nnc_cmd(CCV_NNC_MIN_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}}}, 0); | |||
3291 | ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
3292 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3293 | ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min"); | |||
3294 | } | |||
3295 | ||||
3296 | static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3297 | ||||
3298 | static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = { | |||
3299 | .build = _ccv_cnnp_min_build, | |||
3300 | .copy = _ccv_cnnp_min_copy, | |||
3301 | }; | |||
3302 | ||||
3303 | ccv_cnnp_model_t* ccv_cnnp_min(const char* const name) | |||
3304 | { | |||
3305 | ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t)); | |||
3306 | model_min->super.isa = &ccv_cnnp_min_isa; | |||
3307 | model_min->super.input_size = 2; | |||
3308 | model_min->super.outputs = &model_min->output; | |||
3309 | model_min->super.output_size = 1; | |||
3310 | ccv_cnnp_model_copy_name(&model_min->super, name); | |||
3311 | return (ccv_cnnp_model_t*)model_min; | |||
3312 | } | |||
3313 | ||||
3314 | static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3315 | { | |||
3316 | const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super; | |||
3317 | return ccv_cnnp_min(self->super.name); | |||
3318 | } | |||
3319 | ||||
3320 | // MARK - Max Layer | |||
3321 | ||||
3322 | typedef struct { | |||
3323 | ccv_cnnp_model_t super; | |||
3324 | ccv_nnc_tensor_symbol_t output; | |||
3325 | } ccv_cnnp_model_max_t; | |||
3326 | ||||
3327 | static void _ccv_cnnp_max_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3328 | { | |||
3329 | PRINT(CCV_CLI_VERBOSE, "[cnnp_max_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_max_build] -\n"); fflush(stdout); } } while ( 0); | |||
3330 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 3330, __extension__ __PRETTY_FUNCTION__); })); | |||
3331 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3331, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3332 | ccv_nnc_tensor_param_t input_params[2]; | |||
3333 | int i; | |||
3334 | for (i = 0; i < 2; i++) | |||
3335 | input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]); | |||
3336 | ccv_nnc_tensor_param_t output_params; | |||
3337 | const ccv_nnc_cmd_t max = CMD_MAX_FORWARD()ccv_nnc_cmd(CCV_NNC_MAX_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}}}, 0); | |||
3338 | ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1); | |||
3339 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0); | |||
3340 | ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max"); | |||
3341 | } | |||
3342 | ||||
3343 | static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3344 | ||||
3345 | static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = { | |||
3346 | .build = _ccv_cnnp_max_build, | |||
3347 | .copy = _ccv_cnnp_max_copy, | |||
3348 | }; | |||
3349 | ||||
3350 | ccv_cnnp_model_t* ccv_cnnp_max(const char* const name) | |||
3351 | { | |||
3352 | ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t)); | |||
3353 | model_max->super.isa = &ccv_cnnp_max_isa; | |||
3354 | model_max->super.input_size = 2; | |||
3355 | model_max->super.outputs = &model_max->output; | |||
3356 | model_max->super.output_size = 1; | |||
3357 | ccv_cnnp_model_copy_name(&model_max->super, name); | |||
3358 | return (ccv_cnnp_model_t*)model_max; | |||
3359 | } | |||
3360 | ||||
3361 | static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3362 | { | |||
3363 | const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super; | |||
3364 | return ccv_cnnp_max(self->super.name); | |||
3365 | } | |||
3366 | ||||
3367 | // MARK - LSTM Layer | |||
3368 | ||||
3369 | typedef struct { | |||
3370 | ccv_cnnp_model_t super; | |||
3371 | int masked; | |||
3372 | ccv_nnc_tensor_symbol_t output; | |||
3373 | ccv_nnc_tensor_symbol_t weights; | |||
3374 | ccv_nnc_tensor_symbol_t reserves; | |||
3375 | ccv_nnc_cmd_param_t params; | |||
3376 | ccv_nnc_graph_exec_symbol_t lstm; | |||
3377 | } ccv_cnnp_model_lstm_t; | |||
3378 | ||||
3379 | static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias) | |||
3380 | { | |||
3381 | const int D = !!bidirectional + 1; | |||
3382 | if (hidden_size == proj_size) | |||
3383 | return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D; | |||
3384 | else | |||
3385 | return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (proj_size * 4 * D + proj_size * 4) + (proj_size * 4 + input_size * 4) + num_layers * proj_size) * D; | |||
3386 | } | |||
3387 | ||||
3388 | static void _ccv_cnnp_lstm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3389 | { | |||
3390 | ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super; | |||
3391 | PRINT(CCV_CLI_VERBOSE, "[cnnp_lstm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_lstm_build] -\n"); fflush(stdout); } } while (0); | |||
3392 | assert(input_size == self->super.input_size)((void) sizeof ((input_size == self->super.input_size) ? 1 : 0), __extension__ ({ if (input_size == self->super.input_size ) ; else __assert_fail ("input_size == self->super.input_size" , "ccv_cnnp_model_addons.c", 3392, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3393 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3393, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3394 | const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size; | |||
3395 | ccv_nnc_tensor_param_t input_params[5]; | |||
3396 | input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3397 | if (input_size == 2) | |||
3398 | input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]); | |||
3399 | input_params[4] = input_params[0]; | |||
3400 | memset(input_params[4].dim, 0, sizeof(input_params[4].dim)); | |||
3401 | const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim); | |||
3402 | const int feature_count = input_params[0].dim[x_nd - 1]; | |||
3403 | input_params[4].dim[0] = _ccv_cnnp_lstm_weight_dim(self->params.rnn.bidirectional, self->params.rnn.num_layers, feature_count, self->params.rnn.hidden_size, proj_size, self->params.rnn.bias); | |||
3404 | input_params[4].dim[1] = self->params.rnn.hidden_size; | |||
3405 | const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0); | |||
3406 | ccv_nnc_tensor_param_t output_params[4]; | |||
3407 | ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4); | |||
3408 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
3409 | if (!self->weights.graph) | |||
3410 | self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights"); | |||
3411 | if (!self->reserves.graph) | |||
3412 | self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves"); | |||
3413 | const ccv_nnc_tensor_symbol_t mask = input_size == 2 ? inputs[1] : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
3414 | self->lstm = ccv_nnc_graph_exec_symbol_new(graph, lstm, TENSOR_SYMBOL_LIST(inputs[0], mask, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, self->weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], mask, (const ccv_nnc_tensor_symbol_t ){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t ){.d = CCV_NNC_NO_TENSOR_SYMBOL}, self->weights}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 - 1), TENSOR_SYMBOL_LIST(outputs[0], NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, self->reserves)(const ccv_nnc_tensor_symbol_t []){outputs[0], (const ccv_nnc_tensor_symbol_t ){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t ){.d = CCV_NNC_NO_TENSOR_SYMBOL}, self->reserves}, (1 +1 + 1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "lstm"); | |||
3415 | } | |||
3416 | ||||
3417 | static void _ccv_cnnp_lstm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
3418 | { | |||
3419 | ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super; | |||
3420 | if (self->weights.graph) | |||
3421 | { | |||
3422 | const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size); | |||
3423 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-stdv, stdv)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-stdv, stdv}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
3424 | } | |||
3425 | } | |||
3426 | ||||
3427 | static void _ccv_cnnp_lstm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
3428 | { | |||
3429 | ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super; | |||
3430 | if (self->weights.graph) | |||
3431 | add_to_array(parameters, self->weights, is_trainable); | |||
3432 | } | |||
3433 | ||||
3434 | static void _ccv_cnnp_lstm_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context) | |||
3435 | { | |||
3436 | ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super; | |||
3437 | if (self->lstm.graph) | |||
3438 | { | |||
3439 | self->params.rnn.is_test = is_test; | |||
3440 | updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint); | |||
3441 | } | |||
3442 | } | |||
3443 | ||||
3444 | static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3445 | ||||
3446 | static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = { | |||
3447 | .build = _ccv_cnnp_lstm_build, | |||
3448 | .init_states = _ccv_cnnp_lstm_init_states, | |||
3449 | .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter, | |||
3450 | .copy = _ccv_cnnp_lstm_copy, | |||
3451 | .set_is_test = _ccv_cnnp_lstm_set_is_test, | |||
3452 | }; | |||
3453 | ||||
3454 | ccv_cnnp_model_t* ccv_cnnp_lstm(const int masked, const int hidden_size, const int proj_size, const int num_layers, const int bias, const int batch_first, const int bidirectional, const float dropout, const int is_trainable, const char* const name) | |||
3455 | { | |||
3456 | ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t)); | |||
3457 | model_lstm->super.isa = &ccv_cnnp_lstm_isa; | |||
3458 | model_lstm->super.input_size = masked ? 2 : 1; | |||
3459 | model_lstm->super.outputs = &model_lstm->output; | |||
3460 | model_lstm->super.output_size = 1; | |||
3461 | model_lstm->super.is_trainable = is_trainable; | |||
3462 | ccv_cnnp_model_copy_name(&model_lstm->super, name); | |||
3463 | model_lstm->masked = masked; | |||
3464 | model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
3465 | model_lstm->weights.graph = 0; | |||
3466 | model_lstm->params.rnn.hidden_size = hidden_size; | |||
3467 | model_lstm->params.rnn.proj_size = proj_size; | |||
3468 | model_lstm->params.rnn.num_layers = num_layers; | |||
3469 | model_lstm->params.rnn.bias = bias; | |||
3470 | model_lstm->params.rnn.batch_first = batch_first; | |||
3471 | model_lstm->params.rnn.bidirectional = bidirectional; | |||
3472 | model_lstm->params.rnn.dropout = dropout; | |||
3473 | return (ccv_cnnp_model_t*)model_lstm; | |||
3474 | } | |||
3475 | ||||
3476 | static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3477 | { | |||
3478 | const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super; | |||
3479 | return ccv_cnnp_lstm(self->masked, self->params.rnn.hidden_size, self->params.rnn.proj_size, self->params.rnn.num_layers, self->params.rnn.bias, self->params.rnn.batch_first, self->params.rnn.bidirectional, self->params.rnn.dropout, self->super.is_trainable, self->super.name); | |||
3480 | } | |||
3481 | ||||
3482 | /// MARK - Datatype conversion layer. | |||
3483 | ||||
3484 | typedef struct { | |||
3485 | ccv_cnnp_model_t super; | |||
3486 | ccv_nnc_tensor_symbol_t output; | |||
3487 | int datatype; | |||
3488 | int ref_to_last; | |||
3489 | } ccv_cnnp_model_datatype_conversion_t; | |||
3490 | ||||
3491 | static void _ccv_cnnp_datatype_conversion_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3492 | { | |||
3493 | ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super; | |||
3494 | PRINT(CCV_CLI_VERBOSE, "[cnnp_datatype_conversion_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_datatype_conversion_build] -\n"); fflush(stdout ); } } while (0); | |||
3495 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3496 | if (self->ref_to_last) | |||
3497 | { | |||
3498 | assert(input_size > 1)((void) sizeof ((input_size > 1) ? 1 : 0), __extension__ ( { if (input_size > 1) ; else __assert_fail ("input_size > 1" , "ccv_cnnp_model_addons.c", 3498, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3499 | const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]); | |||
3500 | params.datatype = last_params.datatype; | |||
3501 | } else | |||
3502 | params.datatype = self->datatype; | |||
3503 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3503, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3504 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
3505 | ccv_nnc_graph_exec_symbol_new(graph, CMD_DATATYPE_CONVERSION_FORWARD()ccv_nnc_cmd(CCV_NNC_DATATYPE_CONVERSION_FORWARD, 0, ccv_nnc_cmd_auto , 0), inputs, output_size, outputs, output_size, 0); | |||
3506 | } | |||
3507 | ||||
3508 | static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3509 | ||||
3510 | static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = { | |||
3511 | .build = _ccv_cnnp_datatype_conversion_build, | |||
3512 | .copy = _ccv_cnnp_datatype_conversion_copy, | |||
3513 | }; | |||
3514 | ||||
3515 | ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name) | |||
3516 | { | |||
3517 | ccv_cnnp_model_datatype_conversion_t* const model_datatype_conversion = (ccv_cnnp_model_datatype_conversion_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_datatype_conversion_t)); | |||
3518 | model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa; | |||
3519 | model_datatype_conversion->super.input_size = 0; | |||
3520 | model_datatype_conversion->super.outputs = &model_datatype_conversion->output; | |||
3521 | model_datatype_conversion->super.output_size = 1; | |||
3522 | model_datatype_conversion->datatype = datatype; | |||
3523 | model_datatype_conversion->ref_to_last = ref_to_last; | |||
3524 | ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name); | |||
3525 | return (ccv_cnnp_model_t*)model_datatype_conversion; | |||
3526 | } | |||
3527 | ||||
3528 | static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3529 | { | |||
3530 | ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super; | |||
3531 | return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name); | |||
3532 | } | |||
3533 | ||||
3534 | /// MARK - Clamp layer. | |||
3535 | ||||
3536 | typedef struct { | |||
3537 | ccv_cnnp_model_t super; | |||
3538 | ccv_nnc_tensor_symbol_t output; | |||
3539 | float min; | |||
3540 | float max; | |||
3541 | } ccv_cnnp_model_clamp_t; | |||
3542 | ||||
3543 | static void _ccv_cnnp_clamp_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3544 | { | |||
3545 | ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super; | |||
3546 | PRINT(CCV_CLI_VERBOSE, "[cnnp_clamp_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_clamp_build] -\n"); fflush(stdout); } } while (0); | |||
3547 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3548 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3548, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3549 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
3550 | ccv_nnc_graph_exec_symbol_new(graph, CMD_CLAMP_FORWARD(self->min, self->max)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){. size={.dim={1,1,1}},.clamp={.min=self->min,.max=self->max }}, 0), inputs, output_size, outputs, output_size, 0); | |||
3551 | } | |||
3552 | ||||
3553 | static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context); | |||
3554 | ||||
3555 | static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = { | |||
3556 | .build = _ccv_cnnp_clamp_build, | |||
3557 | .copy = _ccv_cnnp_clamp_copy, | |||
3558 | }; | |||
3559 | ||||
3560 | ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name) | |||
3561 | { | |||
3562 | ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t)); | |||
3563 | model_clamp->super.isa = &ccv_cnnp_clamp_isa; | |||
3564 | model_clamp->super.input_size = 0; | |||
3565 | model_clamp->super.outputs = &model_clamp->output; | |||
3566 | model_clamp->super.output_size = 1; | |||
3567 | model_clamp->min = min; | |||
3568 | model_clamp->max = max; | |||
3569 | ccv_cnnp_model_copy_name(&model_clamp->super, name); | |||
3570 | return (ccv_cnnp_model_t*)model_clamp; | |||
3571 | } | |||
3572 | ||||
3573 | static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3574 | { | |||
3575 | ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super; | |||
3576 | return ccv_cnnp_clamp(self->min, self->max, self->super.name); | |||
3577 | } | |||
3578 | ||||
3579 | // MARK - Parameter Layer | |||
3580 | ||||
3581 | typedef struct { | |||
3582 | ccv_cnnp_model_t super; | |||
3583 | float init_bound; | |||
3584 | ccv_nnc_tensor_symbol_t weights; | |||
3585 | ccv_nnc_tensor_param_t weights_params; | |||
3586 | ccv_nnc_tensor_symbol_t output; | |||
3587 | } ccv_cnnp_model_parameter_t; | |||
3588 | ||||
3589 | static void _ccv_cnnp_parameter_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3590 | { | |||
3591 | PRINT(CCV_CLI_VERBOSE, "[cnnp_parameter_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_parameter_build] -\n"); fflush(stdout); } } while (0); | |||
3592 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3592, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3593 | ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super; | |||
3594 | if (!self->weights.graph) | |||
3595 | self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights"); | |||
3596 | assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__ ({ if (self->weights.graph == graph) ; else __assert_fail ("self->weights.graph == graph", "ccv_cnnp_model_addons.c" , 3596, __extension__ __PRETTY_FUNCTION__); })); | |||
3597 | outputs[0] = self->weights; | |||
3598 | } | |||
3599 | ||||
3600 | static void _ccv_cnnp_parameter_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
3601 | { | |||
3602 | ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super; | |||
3603 | if (self->init_bound > 0) | |||
3604 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-self->init_bound, self->init_bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-self->init_bound, self-> init_bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
3605 | else | |||
3606 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
3607 | } | |||
3608 | ||||
3609 | static void _ccv_cnnp_parameter_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
3610 | { | |||
3611 | ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super; | |||
3612 | add_to_array(parameters, self->weights, is_trainable); | |||
3613 | } | |||
3614 | ||||
3615 | static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3616 | ||||
3617 | static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = { | |||
3618 | .build = _ccv_cnnp_parameter_build, | |||
3619 | .init_states = _ccv_cnnp_parameter_init_states, | |||
3620 | .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter, | |||
3621 | .copy = _ccv_cnnp_parameter_copy, | |||
3622 | }; | |||
3623 | ||||
3624 | ccv_cnnp_model_t* ccv_cnnp_parameter(const ccv_nnc_tensor_param_t params, const float init_bound, const int is_trainable, const char* const name) | |||
3625 | { | |||
3626 | ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t)); | |||
3627 | model_parameter->super.isa = &ccv_cnnp_parameter_isa; | |||
3628 | model_parameter->super.input_size = 0; | |||
3629 | model_parameter->super.outputs = &model_parameter->output; | |||
3630 | model_parameter->super.output_size = 1; | |||
3631 | model_parameter->super.is_trainable = is_trainable; | |||
3632 | ccv_cnnp_model_copy_name(&model_parameter->super, name); | |||
3633 | model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
3634 | model_parameter->weights.graph = 0; | |||
3635 | model_parameter->weights_params = params; | |||
3636 | return (ccv_cnnp_model_t*)model_parameter; | |||
3637 | } | |||
3638 | ||||
3639 | static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3640 | { | |||
3641 | const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super; | |||
3642 | return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name); | |||
3643 | } | |||
3644 | ||||
3645 | // MARK - Scalar Layer | |||
3646 | ||||
3647 | typedef struct { | |||
3648 | ccv_cnnp_model_t super; | |||
3649 | int type; | |||
3650 | int format; | |||
3651 | int datatype; | |||
3652 | float value; | |||
3653 | ccv_nnc_tensor_symbol_t output; | |||
3654 | } ccv_cnnp_model_scalar_t; | |||
3655 | ||||
3656 | static void _ccv_cnnp_scalar_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3657 | { | |||
3658 | PRINT(CCV_CLI_VERBOSE, "[cnnp_scalar_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_scalar_build] -\n"); fflush(stdout); } } while (0); | |||
3659 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3659, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3660 | ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super; | |||
3661 | ccv_nnc_tensor_param_t params = { | |||
3662 | .type = self->type, | |||
3663 | .format = self->format, | |||
3664 | .datatype = self->datatype, | |||
3665 | .dim = { | |||
3666 | 1 | |||
3667 | } | |||
3668 | }; | |||
3669 | if (input_size > 0) | |||
3670 | { | |||
3671 | ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3672 | params.type = input_params.type; | |||
3673 | params.format = input_params.format; | |||
3674 | params.datatype = input_params.datatype; | |||
3675 | } | |||
3676 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
3677 | ccv_nnc_graph_exec_symbol_new(graph, CMD_SET_FORWARD(self->value)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={self->value,}}}, 0), 0, 0, outputs, 1, 0); | |||
3678 | } | |||
3679 | ||||
3680 | static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3681 | ||||
3682 | static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = { | |||
3683 | .build = _ccv_cnnp_scalar_build, | |||
3684 | .copy = _ccv_cnnp_scalar_copy, | |||
3685 | }; | |||
3686 | ||||
3687 | ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name) | |||
3688 | { | |||
3689 | ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t)); | |||
3690 | model_scalar->super.isa = &ccv_cnnp_scalar_isa; | |||
3691 | model_scalar->super.input_size = 0; | |||
3692 | model_scalar->super.outputs = &model_scalar->output; | |||
3693 | model_scalar->super.output_size = 1; | |||
3694 | ccv_cnnp_model_copy_name(&model_scalar->super, name); | |||
3695 | model_scalar->type = type; | |||
3696 | model_scalar->format = format; | |||
3697 | model_scalar->datatype = datatype; | |||
3698 | model_scalar->value = value; | |||
3699 | return (ccv_cnnp_model_t*)model_scalar; | |||
3700 | } | |||
3701 | ||||
3702 | static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3703 | { | |||
3704 | const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super; | |||
3705 | return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name); | |||
3706 | } | |||
3707 | ||||
3708 | // MARK - Variable Layer | |||
3709 | ||||
3710 | typedef struct { | |||
3711 | ccv_cnnp_model_t super; | |||
3712 | ccv_nnc_tensor_param_t params; | |||
3713 | ccv_nnc_tensor_symbol_t output; | |||
3714 | } ccv_cnnp_model_variable_t; | |||
3715 | ||||
3716 | static void _ccv_cnnp_variable_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3717 | { | |||
3718 | PRINT(CCV_CLI_VERBOSE, "[cnnp_variable_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_variable_build] -\n"); fflush(stdout); } } while (0); | |||
3719 | assert(input_size == 0)((void) sizeof ((input_size == 0) ? 1 : 0), __extension__ ({ if (input_size == 0) ; else __assert_fail ("input_size == 0", "ccv_cnnp_model_addons.c" , 3719, __extension__ __PRETTY_FUNCTION__); })); | |||
3720 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3720, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3721 | ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super; | |||
3722 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0); | |||
3723 | } | |||
3724 | ||||
3725 | static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3726 | ||||
3727 | static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = { | |||
3728 | .build = _ccv_cnnp_variable_build, | |||
3729 | .copy = _ccv_cnnp_variable_copy, | |||
3730 | }; | |||
3731 | ||||
3732 | ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name) | |||
3733 | { | |||
3734 | ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t)); | |||
3735 | model_variable->super.isa = &ccv_cnnp_variable_isa; | |||
3736 | model_variable->super.input_size = 0; | |||
3737 | model_variable->super.outputs = &model_variable->output; | |||
3738 | model_variable->super.output_size = 1; | |||
3739 | ccv_cnnp_model_copy_name(&model_variable->super, name); | |||
3740 | model_variable->params = params; | |||
3741 | return (ccv_cnnp_model_t*)model_variable; | |||
3742 | } | |||
3743 | ||||
3744 | static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3745 | { | |||
3746 | const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super; | |||
3747 | return ccv_cnnp_variable(self->params, self->super.name); | |||
3748 | } | |||
3749 | ||||
3750 | // MARK - Move Layer | |||
3751 | ||||
3752 | typedef struct { | |||
3753 | ccv_cnnp_model_t super; | |||
3754 | ccv_nnc_tensor_symbol_t output; | |||
3755 | } ccv_cnnp_model_move_t; | |||
3756 | ||||
3757 | static void _ccv_cnnp_move_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3758 | { | |||
3759 | PRINT(CCV_CLI_VERBOSE, "[cnnp_move_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_move_build] -\n"); fflush(stdout); } } while (0); | |||
3760 | assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if (input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c" , 3760, __extension__ __PRETTY_FUNCTION__); })); | |||
3761 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3761, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3762 | outputs[0] = inputs[1]; | |||
3763 | ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), inputs, 1, outputs, 1, "move"); | |||
3764 | } | |||
3765 | ||||
3766 | static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3767 | ||||
3768 | static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = { | |||
3769 | .build = _ccv_cnnp_move_build, | |||
3770 | .copy = _ccv_cnnp_move_copy, | |||
3771 | }; | |||
3772 | ||||
3773 | ccv_cnnp_model_t* ccv_cnnp_move(const char* const name) | |||
3774 | { | |||
3775 | ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t)); | |||
3776 | model_move->super.isa = &ccv_cnnp_move_isa; | |||
3777 | model_move->super.input_size = 2; | |||
3778 | model_move->super.outputs = &model_move->output; | |||
3779 | model_move->super.output_size = 1; | |||
3780 | ccv_cnnp_model_copy_name(&model_move->super, name); | |||
3781 | return (ccv_cnnp_model_t*)model_move; | |||
3782 | } | |||
3783 | ||||
3784 | static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3785 | { | |||
3786 | const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super; | |||
3787 | return ccv_cnnp_move(self->super.name); | |||
3788 | } | |||
3789 | ||||
3790 | // MARK - "Making" Contiguous Layer | |||
3791 | ||||
3792 | typedef struct { | |||
3793 | ccv_cnnp_model_t super; | |||
3794 | ccv_nnc_tensor_symbol_t output; | |||
3795 | } ccv_cnnp_model_contiguous_t; | |||
3796 | ||||
3797 | static void _ccv_cnnp_contiguous_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3798 | { | |||
3799 | PRINT(CCV_CLI_VERBOSE, "[cnnp_contiguous_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_contiguous_build] -\n"); fflush(stdout); } } while (0); | |||
3800 | assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if (input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c" , 3800, __extension__ __PRETTY_FUNCTION__); })); | |||
3801 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3801, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3802 | ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3803 | ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]); | |||
3804 | if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward. | |||
3805 | { | |||
3806 | outputs[0] = inputs[0]; | |||
3807 | return; | |||
3808 | } | |||
3809 | // Otherwise, we need to check its stride to know if it is contiguous. | |||
3810 | int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)]; | |||
3811 | ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride); | |||
3812 | // We identify permute by checking if the stride is not in descending order. | |||
3813 | // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly. | |||
3814 | if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim)) | |||
3815 | { | |||
3816 | outputs[0] = inputs[0]; | |||
3817 | return; | |||
3818 | } | |||
3819 | outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0); | |||
3820 | ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto , 0), inputs, 1, outputs, 1, "contiguous"); | |||
3821 | ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT); | |||
3822 | } | |||
3823 | ||||
3824 | static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3825 | ||||
3826 | static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = { | |||
3827 | .build = _ccv_cnnp_contiguous_build, | |||
3828 | .copy = _ccv_cnnp_contiguous_copy, | |||
3829 | }; | |||
3830 | ||||
3831 | ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name) | |||
3832 | { | |||
3833 | ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t)); | |||
3834 | model_contiguous->super.isa = &ccv_cnnp_contiguous_isa; | |||
3835 | model_contiguous->super.input_size = 1; | |||
3836 | model_contiguous->super.outputs = &model_contiguous->output; | |||
3837 | model_contiguous->super.output_size = 1; | |||
3838 | ccv_cnnp_model_copy_name(&model_contiguous->super, name); | |||
3839 | return (ccv_cnnp_model_t*)model_contiguous; | |||
3840 | } | |||
3841 | ||||
3842 | static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3843 | { | |||
3844 | const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super; | |||
3845 | return ccv_cnnp_contiguous(self->super.name); | |||
3846 | } | |||
3847 | ||||
3848 | // MARK - Scaled-Dot Product Attention Layer | |||
3849 | ||||
3850 | typedef struct { | |||
3851 | ccv_cnnp_model_t super; | |||
3852 | ccv_nnc_tensor_symbol_t output; | |||
3853 | ccv_nnc_tensor_symbol_t weights; | |||
3854 | ccv_nnc_tensor_symbol_t bias; | |||
3855 | float scale; | |||
3856 | int is_causal; | |||
3857 | int has_attn_mask; | |||
3858 | int upcast; | |||
3859 | int fused_unify_head_weights; | |||
3860 | int no_bias; | |||
3861 | } ccv_cnnp_model_scaled_dot_product_attention_t; | |||
3862 | ||||
3863 | static void _ccv_cnnp_scaled_dot_product_attention_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size) | |||
3864 | { | |||
3865 | PRINT(CCV_CLI_VERBOSE, "[cnnp_scaled_dot_product_attention_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("[cnnp_scaled_dot_product_attention_build] -\n"); fflush (stdout); } } while (0); | |||
3866 | assert(input_size == 3 || input_size == 4)((void) sizeof ((input_size == 3 || input_size == 4) ? 1 : 0) , __extension__ ({ if (input_size == 3 || input_size == 4) ; else __assert_fail ("input_size == 3 || input_size == 4", "ccv_cnnp_model_addons.c" , 3866, __extension__ __PRETTY_FUNCTION__); })); | |||
3867 | assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({ if (output_size == 1) ; else __assert_fail ("output_size == 1" , "ccv_cnnp_model_addons.c", 3867, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3868 | ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super; | |||
3869 | const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]); | |||
3870 | const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]); | |||
3871 | const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]); | |||
3872 | const int v_nd = ccv_nnc_tensor_nd(v_params.dim); | |||
3873 | assert(v_nd == 3 || v_nd == 4)((void) sizeof ((v_nd == 3 || v_nd == 4) ? 1 : 0), __extension__ ({ if (v_nd == 3 || v_nd == 4) ; else __assert_fail ("v_nd == 3 || v_nd == 4" , "ccv_cnnp_model_addons.c", 3873, __extension__ __PRETTY_FUNCTION__ ); })); | |||
3874 | const int hEv = (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1]; | |||
3875 | ccv_nnc_tensor_param_t weights_params = q_params; | |||
3876 | memset(weights_params.dim, 0, sizeof(weights_params.dim)); | |||
3877 | weights_params.dim[0] = hEv; | |||
3878 | weights_params.dim[1] = hEv; | |||
3879 | ccv_nnc_tensor_param_t bias_params = q_params; | |||
3880 | memset(bias_params.dim, 0, sizeof(bias_params.dim)); | |||
3881 | bias_params.dim[0] = hEv; | |||
3882 | ccv_nnc_cmd_t cmd = {0}; | |||
3883 | cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD; | |||
3884 | cmd.info.scaled_dot_product_attention.scale = self->scale; | |||
3885 | cmd.info.scaled_dot_product_attention.is_causal = self->is_causal; | |||
3886 | cmd.info.scaled_dot_product_attention.upcast = self->upcast; | |||
3887 | ccv_nnc_tensor_param_t output_params[3]; | |||
3888 | ccv_nnc_tensor_symbol_t output; | |||
3889 | ccv_nnc_tensor_symbol_t saved_softmax; | |||
3890 | ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
3891 | ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
3892 | ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
3893 | ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
3894 | if (self->has_attn_mask) | |||
3895 | attn_mask = inputs[3]; | |||
3896 | if (self->fused_unify_head_weights) | |||
3897 | { | |||
3898 | if (!self->weights.graph) | |||
3899 | self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights"); | |||
3900 | weights = self->weights; | |||
3901 | if (!self->no_bias) | |||
3902 | { | |||
3903 | if (!self->bias.graph) | |||
3904 | self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias"); | |||
3905 | bias = self->bias; | |||
3906 | } | |||
3907 | ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){ | |||
3908 | q_params, | |||
3909 | k_params, | |||
3910 | v_params, | |||
3911 | (ccv_nnc_tensor_param_t){}, | |||
3912 | weights_params, | |||
3913 | bias_params, | |||
3914 | }, 6, ccv_nnc_no_hint, output_params, 3); | |||
3915 | output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
3916 | saved_softmax = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0); | |||
3917 | saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0); | |||
3918 | } else { | |||
3919 | ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){ | |||
3920 | q_params, | |||
3921 | k_params, | |||
3922 | v_params, | |||
3923 | }, 3, ccv_nnc_no_hint, output_params, 2); | |||
3924 | output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0); | |||
3925 | saved_softmax = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0); | |||
3926 | } | |||
3927 | ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], attn_mask, weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs [2], attn_mask, weights, bias}, (1 +1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax, saved_v_proj }, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 + 0 +0 +0 +0 -1), "scaled_dot_product_attention"); | |||
3928 | outputs[0] = output; | |||
3929 | } | |||
3930 | ||||
3931 | static void _ccv_cnnp_scaled_dot_product_attention_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context) | |||
3932 | { | |||
3933 | ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super; | |||
3934 | if (self->weights.graph) | |||
3935 | { | |||
3936 | assert(self->fused_unify_head_weights)((void) sizeof ((self->fused_unify_head_weights) ? 1 : 0), __extension__ ({ if (self->fused_unify_head_weights) ; else __assert_fail ("self->fused_unify_head_weights", "ccv_cnnp_model_addons.c" , 3936, __extension__ __PRETTY_FUNCTION__); })); | |||
3937 | const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights); | |||
3938 | const int c = weight_params.dim[1]; | |||
3939 | const float std = sqrtf(2) / sqrtf(c); | |||
3940 | const float bound = sqrtf(3) * std; | |||
3941 | initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t ){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights); | |||
3942 | if (self->bias.graph) | |||
3943 | initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias); | |||
3944 | } | |||
3945 | } | |||
3946 | ||||
3947 | static void _ccv_cnnp_scaled_dot_product_attention_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable) | |||
3948 | { | |||
3949 | ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super; | |||
3950 | if (self->weights.graph) | |||
3951 | { | |||
3952 | assert(self->fused_unify_head_weights)((void) sizeof ((self->fused_unify_head_weights) ? 1 : 0), __extension__ ({ if (self->fused_unify_head_weights) ; else __assert_fail ("self->fused_unify_head_weights", "ccv_cnnp_model_addons.c" , 3952, __extension__ __PRETTY_FUNCTION__); })); | |||
3953 | add_to_array(parameters, self->weights, is_trainable); | |||
3954 | if (self->bias.graph) | |||
3955 | add_to_array(parameters, self->bias, is_trainable); | |||
3956 | } | |||
3957 | } | |||
3958 | ||||
3959 | static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context); | |||
3960 | ||||
3961 | static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = { | |||
3962 | .build = _ccv_cnnp_scaled_dot_product_attention_build, | |||
3963 | .copy = _ccv_cnnp_scaled_dot_product_attention_copy, | |||
3964 | }; | |||
3965 | ||||
3966 | static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = { | |||
3967 | .build = _ccv_cnnp_scaled_dot_product_attention_build, | |||
3968 | .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states, | |||
3969 | .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter, | |||
3970 | .copy = _ccv_cnnp_scaled_dot_product_attention_copy, | |||
3971 | }; | |||
3972 | ||||
3973 | ccv_cnnp_model_t* ccv_cnnp_scaled_dot_product_attention(const float scale, const int is_causal, const int has_attn_mask, const int upcast, const int fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name) | |||
3974 | { | |||
3975 | ccv_cnnp_model_scaled_dot_product_attention_t* const model_scaled_dot_product_attention = (ccv_cnnp_model_scaled_dot_product_attention_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scaled_dot_product_attention_t)); | |||
3976 | model_scaled_dot_product_attention->super.isa = fused_unify_head_weights ? &ccv_cnnp_scaled_dot_product_attention_fused_isa : &ccv_cnnp_scaled_dot_product_attention_isa; | |||
3977 | model_scaled_dot_product_attention->super.input_size = has_attn_mask ? 4 : 3; | |||
3978 | model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output; | |||
3979 | model_scaled_dot_product_attention->super.output_size = 1; | |||
3980 | model_scaled_dot_product_attention->super.is_trainable = is_trainable; | |||
3981 | ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name); | |||
3982 | model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
3983 | model_scaled_dot_product_attention->weights.graph = 0; | |||
3984 | model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL; | |||
3985 | model_scaled_dot_product_attention->bias.graph = 0; | |||
3986 | model_scaled_dot_product_attention->scale = scale; | |||
3987 | model_scaled_dot_product_attention->is_causal = is_causal; | |||
3988 | model_scaled_dot_product_attention->has_attn_mask = has_attn_mask; | |||
3989 | model_scaled_dot_product_attention->upcast = upcast; | |||
3990 | model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights; | |||
3991 | model_scaled_dot_product_attention->no_bias = no_bias; | |||
3992 | return (ccv_cnnp_model_t*)model_scaled_dot_product_attention; | |||
3993 | } | |||
3994 | ||||
3995 | static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context) | |||
3996 | { | |||
3997 | const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super; | |||
3998 | return ccv_cnnp_scaled_dot_product_attention(self->scale, self->is_causal, self->has_attn_mask, self->upcast, self->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name); | |||
3999 | } |